Episode 112: Google Antigravity, ChatGPT Mac with WebMCP

AgentStack Daily EP112 — Google Antigravity, ChatGPT Mac with WebMCP, MiniMax-Text-01, and Claude Code CLI

Title: Google Antigravity, ChatGPT Mac with WebMCP, and MiniMax-Text-01's 4M Context Engine

Tagline: Google Antigravity advances agentic development across Antigravity 2.0, Go CLI, and Python SDK. OpenAI updates ChatGPT on Mac with Computer History and WebMCP site tools. MiniMax-Text-01 pushes open-weights reasoning with a 4-million-token context window and Lightning Attention. Claude Code CLI 2.1.267 strengthens terminal agent workflows while Hermes Agent rolls out v2026.9.7. Mercury 2.5 lands diffusion-style reasoning on OpenRouter. Cognition secures a $48B valuation, and OpenAI posts an AI-generated Navier-Stokes Lean proof.

Feed description: Google Antigravity expands its agentic platform across desktop, CLI, and SDK. OpenAI brings Computer History and WebMCP site tools to ChatGPT on Mac. MiniMax-Text-01 delivers open-weights MoE reasoning with 4M context. Claude Code CLI 2.1.267 and Hermes Agent v2026.9.7 update the agent stack. Mercury 2.5 brings diffusion reasoning to OpenRouter. Plus Cognition's $48B round, 1Password's 21% Codex lift, Mistral's €3B raise, and an AI Navier-Stokes proof in Lean.

Story Slate

Agent Stack Release Readout: Claude Code CLI 2.1.267; Hermes Agent v2026.9.7 Primary release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.9.7 Claude Code CLI shipped version 2.1.267 on npm, refining terminal-native agent execution, subagent task isolation, and persistent memory across developer workflows. The update improves large repository exploration, tightens file editing boundaries, and smooths git branching integration during interactive coding sessions. Meanwhile, Hermes Agent rolled out v2026.9.7 with focused maintenance refinements and candidate-state stability improvements for long-running gateways. The release sharpens how background services recover from transient disconnections without dropping active session state. Technical depth angle: Claude Code CLI strengthens terminal-based autonomous coding by enhancing subagent delegation, reducing token overhead during repository indexing, and improving file diff application across dirty working directories. Hermes Agent v2026.9.7 complements this with isolated candidate-state execution, ensuring long-running agent gateways preserve session state through maintenance cycles. Actionability angle: Update Claude Code CLI via npm to gain smoother subagent isolation and cleaner diff resolution on complex repos. For Hermes Agent deployments, v2026.9.7 is a recommended upgrade to secure background gateway stability. The implications land across Antigravity, Codex, Claude Code, and Hermes stacks alike. Listener hook: Your command-line coding agent just got better at handling big codebases without stepping on your git branches.
Google Antigravity: The Agent-First Platform with Antigravity 2.0, Go CLI, and Python SDK Google has organized its developer-facing agent tooling into the Google Antigravity platform, spanning three tightly integrated surfaces: the Antigravity 2.0 desktop environment, a high-speed terminal CLI built in Go, and the programmatic Python SDK. Antigravity treats agents not as disposable prompt wrappers, but as persistent, autonomous collaborators capable of running parallel background tasks, leasing subagents, and producing structured interactive artifacts. The desktop application pairs conversational code exploration with an auxiliary pane managing live file modifications, terminal sessions, and visual diffs. Technical depth angle: Antigravity decouples planning from execution: the main agent can dispatch specialized subagents to handle bounded research or testing tasks while retaining overall control. The Go CLI (agy) provides sub-millisecond local startup and command dispatch, while the Python SDK (google-antigravity) exposes programmatic agent leasing, custom MCP tool registration, and structured event streaming. Actionability angle: Adopt the Antigravity CLI for fast shell-based script automation and headless agent runs. Teams building complex multi-agent pipelines should evaluate the Python SDK for orchestration. Integrate custom project rules and skills to keep autonomous coding aligned with team standards. Listener hook: Google is betting that the future of coding is agent-first, giving you a desktop command center, a fast CLI, and a Python SDK to run them all.
ChatGPT for Mac Rebuilds into Unified Workspace with Computer History and WebMCP Site Tools OpenAI has overhauled its macOS desktop application, merging Chat, Work, and Codex into a unified native client designed for deep workflow automation. The centerpiece is Computer History, an opt-in, privacy-preserving feature that builds a searchable local timeline of user activity across supported applications and websites without capturing screenshots or audio. In addition, the built-in browser now supports WebMCP site tools, allowing the assistant to trigger structured actions—such as commenting in document editors or navigating dashboards—directly inside web applications. Technical depth angle: The unified Mac client bridges desktop context and browser automation. Computer History indexes application-level events into a queryable local store that the model can reference for temporal context. WebMCP extends standard Model Context Protocol concepts to web pages, letting websites expose declarative action schemas that the desktop agent invokes natively. Actionability angle: Opt into Computer History if you want the desktop agent to reference recent application context without manual copying and pasting. Explore WebMCP site tools to automate repetitive web workflows directly from your desktop assistant. Ensure organizational privacy policies align with local activity indexing before enabling broad app permissions. Listener hook: The ChatGPT app on your Mac can now reference what you worked on earlier today and drive web tools directly through your browser.
MiniMax-Text-01 Delivers 4M Context Open-Weights Architecture with Lightning Attention MoE Chinese AI lab MiniMax has expanded its frontier open-weight footprint with MiniMax-Text-01, a 456-billion parameter Mixture-of-Experts architecture activating 45.9 billion parameters per token. Engineered specifically for complex reasoning, autonomous software engineering, and long-horizon agent workflows, the model features a native 4-million-token context window powered by MiniMax's proprietary Lightning Attention mechanism. The linear-complexity attention design avoids quadratic compute growth across millions of tokens, unlocking unprecedented working memory for large codebases and complex multi-agent conversations. Technical depth angle: Lightning Attention replaces full quadratic self-attention with chunked linear recurrence, shrinking KV cache memory footprints and maintaining steady token throughput across million-token sequences. MiniMax-Text-01 couples this with dense router networks that select specialized expert blocks for coding, mathematics, and long-context document synthesis. Actionability angle: Evaluate MiniMax-Text-01 through API endpoints or hosted inference for massive context analysis that would overflow conventional 128K windows. Test the model on long-horizon software engineering benchmarks where entire repositories must fit inside a single agent prompt. Verify that local serving hardware meets the 45.9B active parameter requirements before attempting local deployment. Listener hook: Four million tokens of context in an open-weights mixture-of-experts model changes what an autonomous agent can remember in a single session.
Mercury 2.5 is a diffusion-style reasoning model, now on OpenRouter Inception's Mercury 2.5 has appeared on OpenRouter as a 260,000-token context reasoning model. Mercury 2.5 is what Inception calls a diffusion language model, which generates and refines many tokens in parallel rather than one at a time. The company markets it as the fastest reasoning model it ships. The listing caps output at 4,096 tokens per call. Technical depth angle: Diffusion language models borrow the parallel-refinement idea from image diffusion: instead of predicting the next token in a strict left-to-right chain, the model drafts and then iteratively cleans up many token positions at once. Inception claims that parallelism is what makes Mercury 2.5 faster on reasoning workloads than a sequential model. Actionability angle: For builders, the relevant question is whether the parallel-token approach delivers a real speed edge on long-context reasoning tasks where latency matters. The 260K context window also means very long documents or full chat histories can fit in a single call rather than being chunked and reassembled. Mercury 2.5 is worth a side-by-side test against a sequential reasoning model on the same prompts to see where the speed claim actually shows up. Listener hook: Most language models read left to right, one token at a time. Mercury 2.5 fills in the whole page at once.
Meta's Muse Agent Wants Your Email, Calendar, and Payments Meta launched Muse, a personal AI agent that wants access to email, calendars, payments, and health services — its biggest consumer AI push yet. The bet is simple: people will hand over deeply personal data if an assistant earns its keep by handling it. Reception so far is mixed, with the launch drawing a 505-upvote Hacker News thread where skeptics crowded the comments. What this is really testing is whether Meta, after years of privacy scandals, can get users to trust it as the operator of an always-on personal assistant. Technical depth angle: Muse is a personal AI agent, meaning it acts on your behalf across apps and services rather than just answering questions in a chat box. The mechanism is consent plus action: once granted access to email, calendars, payments, and health data, Muse can read, draft, schedule, and transact for you. The interesting question isn't the model itself — it's the integration surface and what Meta promises, and doesn't, about how that data flows. Actionability angle: For builders, Muse signals that the agent category is now its own design surface, separate from chat, where the permission model is fast becoming the real product. What this means is that consent flows, per-task versus blanket access, will quietly shape which assistants users actually let into their lives. Listener hook: Meta just asked for the keys to your inbox, your bank, and your calendar — and the answer isn't obviously yes.
Cognition's $48B valuation says AI coding won't be a one-tool market Cognition just landed a $48 billion valuation, putting it among the most valuable private AI companies. The number matters less for its size and more for what it tells us about the market: investors clearly don't think AI coding tools will consolidate into a single winner. Cognition's valuation multiple is higher than Cursor's was before SpaceX acquired it. Technical depth angle: The story here is market structure, not a single product mechanism. Cognition sits in a crowded AI coding field, yet private capital is still pricing it at a premium multiple — a signal that buyers expect several durable businesses to coexist rather than one consolidator to swallow the rest. Actionability angle: What this means for builders: the AI coding tool landscape is going to stay crowded, with serious capital backing several serious products, which usually translates into faster feature iteration and more pricing competition. Why it matters: capital is still flowing to alternatives rather than concentrating in a single winner, so the ecosystem isn't about to contract around one vendor. Listener hook: If you've wondered whether AI coding will turn into a one-company market, this valuation says no.
1Password says Codex lifted engineering output 21% 1Password says its engineers have raised productivity by 21% using OpenAI's Codex coding agent, with work going from idea to production while keeping the company's strict security posture intact. The figure comes from a customer story OpenAI published on September 8, framing Codex as a tool the password manager's team now relies on for both customer-facing features and internal tooling. Technical depth angle: The takeaway is workflow-level: Codex is being used inside an organization where every change has to clear a security bar, and 1Password says production-bound code can still pass that bar. Actionability angle: What this means: a 21% lift is a self-reported customer metric, so treat it as one data point rather than a benchmark. For teams with strict quality gates, the interesting question is whether an agent like Codex slots into the existing pipeline without weakening it. Listener hook: A password manager is putting AI-generated code into production under a real security bar, and says it is paying off.
Mistral Closes €3 Billion Series D at €21 Billion Valuation French AI lab Mistral announced on September 8 a €3 billion Series D funding round at a post-money valuation above €21 billion. The figure came directly from Mistral's own blog and quickly drew 822 upvotes on Hacker News. The post itself named no lead investors, gave no breakdown of use of proceeds, and tied the raise to no specific model release or product launch. Technical depth angle: The sourced mechanics are simple and concrete: a €3B Series D raise, a post-money valuation above €21B, announced directly by Mistral on September 8 under the framing 'sovereign open-weight AI to frontier.' The announcement contains no investor names, no compute or headcount commitments, and no product roadmap, so any further mechanism would be invention. Actionability angle: For builders running on open-weight models, this signals continued capital behind a sovereign, downloadable-weight alternative to closed U.S. APIs, but it is a balance-sheet event, not a product change. What this means in practice is that the next Mistral model drop — not the funding round itself — is the signal worth tracking, because capital at this scale changes how aggressively training and distribution can be pushed. Listener hook: A European open-weight AI lab just became one of the most valuable AI companies on paper, and the announcement is telling you exactly what to watch for next.
OpenBMB Ships MiniCPM5-2B, a Sub-3B Open-Weights Model Built to Run On Device OpenBMB released MiniCPM5-2B, a 2.52 billion parameter dense language model with a 131,072 token native context, averaging 53.9 across 34 benchmarks in its model card — ahead of Qwen3.5-4B at 51.1. The clearest leads show up in tool use, coding agents, and long-context retrieval. Training data, intermediate checkpoints, and final weights all ship under Apache 2.0. GGUF builds start at 1.56 GB and run in llama.cpp, Ollama, MLX, vLLM, and SGLang without a fork. Technical depth angle: The post-training pipeline combines 400B tokens of deep-thinking supervised fine-tuning with RL teachers and on-policy distillation that merges 16 expert models into one checkpoint, letting a small dense model inherit specialist behavior without a mixture-of-experts runtime at inference. Actionability angle: What this means is that a sub-3B model with a 131K context window, tool-use capability, and open training data is now downloadable and runnable locally on a high-memory laptop. Builders studying distillation or continued pre-training can access the intermediate Base, Midtrain, and SFT-only checkpoints alongside the final weights. The standard LlamaForCausalLM architecture means existing vLLM, SGLang, llama.cpp, Ollama, and MLX stacks work without modification. Listener hook: A sub-3B model with a 131K context, tool-use skills, and full open training data just landed — and it fits on a high-memory laptop.
Qualcomm and AWS team up on custom AI inference silicon and 1.6T optical links Qualcomm Technologies announced a multi-generation collaboration with Amazon to build customized AI inference silicon for large-scale data centers, alongside optical connectivity work targeting 1.6T. The deal, announced September 8 from San Diego, positions Qualcomm as a serious player in the custom AI chip market that has so far been dominated by companies like Broadcom helping Google's TPUs. The 1.6T optical piece matters because bandwidth between accelerators is becoming the bottleneck as model sizes and cluster sizes grow. Technical depth angle: Two pieces: custom inference silicon tuned for AWS workloads, and 1.6 terabit optical connectivity to link accelerators. Optical at that speed lets data centers move more data between chips without the power and distance limits of copper. Actionability angle: For builders, this means future AWS instances may pair Qualcomm-tuned inference accelerators with faster inter-chip networking, which could shift price-performance for hosted inference. Worth watching for AWS region rollouts and whether the silicon shows up under a branded instance family. Listener hook: If you run AI workloads on AWS, the silicon your inference runs on may quietly be changing under you.
PsiQuantum Locks In $100M US Award for Quantum Manufacturing PsiQuantum announced on September 8, 2026 that it has finalized a $100 million federal R&D award from the US Department of Commerce. The award, made under the CHIPS and Science Act, will fund PsiQuantum's research and development for manufacturing critical components for its quantum computing systems in the United States. The Palo Alto company said the definitive documentation is now signed. Technical depth angle: The award funds R&D for manufacturing 'critical components' for PsiQuantum's quantum systems on US soil. The press release does not specify which components, so the technical scope remains opaque — what is locked in is the funding mechanism and signed paperwork, not a production plan. Actionability angle: This means the US government has committed a nine-figure R&D line to PsiQuantum's domestic manufacturing roadmap. For builders, the practical implication is supply-chain: PsiQuantum is moving toward US-made critical components, which could affect future hardware-partner programs and procurement timelines once specific components and facilities are named. Listener hook: A $100M federal award just formalized one of America's largest quantum manufacturing pushes.
OpenAI Posts an AI-Generated Navier–Stokes Proof in Lean OpenAI published an AI-generated solution to the Navier–Stokes Millennium Prize Problem on September 8, including a writeup and a formal proof in Lean. The Navier–Stokes equations govern how fluids move, and the Millennium Prize version asks whether those equations always produce smooth, well-behaved solutions — a long-open question carrying a million-dollar prize. Lean is an interactive theorem prover, software that checks each step of a mathematical argument the way a compiler checks code. The post drew more than 1,200 upvotes on Hacker News within hours and links to both the writeup and the machine-checkable proof. Technical depth angle: Lean is an interactive theorem prover — software that checks each line of a mathematical argument step by step, the way a compiler checks code. Calling the result a Lean proof means the argument is mechanically verifiable rather than something a reader has to take on trust. The source material does not describe which specific techniques the AI used to derive the solution. Actionability angle: For mathematicians and AI researchers, this is a marker that automated theorem proving is reaching problems at the Millennium Prize level, not just narrow contest puzzles. For builders working in formal verification, it signals that proof-checking tools are now strong enough to be a real part of serious research workflows rather than a curiosity. Listener hook: One of math's seven Millennium Prize problems just got an AI-written proof — and it's machine-checkable in Lean.
Reducto's r-1 Parses a Whole Page in One Pass at a Penny Reducto released r-1 on September 1, 2026, a single-pass document parsing model that folds OCR, layout detection, table extraction, formatting, and grounding into one full-page read. The company reports a 20% drop in parsing errors compared to its prior multi-stage agentic pipeline, while cutting the per-page cost to a flat one cent — down from three to six cents. By collapsing five pipeline stages into one model pass, r-1 replaces the chained architecture Reducto previously shipped alongside. For teams running high-volume document intake, the change turns a pipeline decision into a single API call with predictable pricing. Technical depth angle: The model collapses five separate document-understanding steps — text recognition, page layout, table structure, formatting, and grounding text back to its source location — into one forward pass over the full page, removing the orchestration cost and accumulated error of chaining specialized models. Actionability angle: For teams processing contracts, invoices, or filings at scale, r-1's flat one-cent pricing and lower error rate change the economics of full-document extraction. Migrating from the older pipeline likely means a simpler integration with fewer disagreement bugs between stages. Watch whether the single-pass approach holds up on noisier scans and multi-column layouts. Listener hook: Reducto's new r-1 model turns document parsing from a five-stage pipeline into a single one-cent-per-page call, with twenty percent fewer errors to debug.

Editorial Mix Check

flagship_products: 4
builder_projects: 5
local_ai: 2
hardware_compute: 2
policy_regulation: 1
research: 1

Model Discovery Check

Inception: Mercury 2.5 (inception) — Newly listed this cycle (verified September 09, 2026). Primary source: https://openrouter.ai/models/inception/mercury-2.5. Availability: API via OpenRouter. params_active: n/a; params_total: n/a; context: 260000 tokens; modality: text. Capabilities: context length 260000; Mercury 2.5 is the fastest reasoning LLM, and the latest diffusion LLM (dLLM) from Inception. Instead of generating tokens sequentially, Mercury 2.5 produces and refines multiple tokens in parallel, achieving faster reasoning. Try now / integration angle: Route a coding-agent session through https://openrouter.ai/models/inception/mercury-2.5 and compare it with the current default. Decision: Selected — new major-provider model not featured on a recent broadcast.
Nex AGI: Nex-N2.5-Mini (free) (nex-agi) — Newly listed this cycle (verified September 09, 2026). Primary source: https://openrouter.ai/models/nex-agi/nex-n2.5-mini:free. Availability: API via OpenRouter. Capabilities: context length 262144; Nex-N2.5 is an agentic model built to turn goals into working, verified outcomes. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.
Nex AGI: Nex-N2.5-Pro (free) (nex-agi) — Newly listed this cycle (verified September 09, 2026). Primary source: https://openrouter.ai/models/nex-agi/nex-n2.5-pro:free. Availability: API via OpenRouter. Capabilities: context length 262144; Nex-N2.5 is an agentic model built to turn goals into working, verified outcomes. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.
DeepSeek: DeepSeek V4 Flash Vision Exp (batch) (deepseek) — Newly listed this cycle (verified September 09, 2026). Primary source: https://openrouter.ai/models/deepseek/deepseek-v4-flash-vision-exp:batch. Availability: API via OpenRouter. Capabilities: context length 1048576; DeepSeek V4 Flash Vision Exp is an experimental vision-enabled version of DeepSeek V4 Flash. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.
Z.ai: GLM 5.3 (batch) (z-ai) — Newly listed this cycle (verified September 09, 2026). Primary source: https://openrouter.ai/models/z-ai/glm-5.3:batch. Availability: API via OpenRouter. Capabilities: context length 1048576; GLM-5.3 is a large-scale reasoning model from Z.ai for software engineering. Try now / integration angle: available for evaluation via the model page above. Decision: Not Selected — variant/duplicate of a model featured on a recent broadcast, or not a major standalone drop.

Local LLM Spotlight

Spark-X2.5-4B (XHToken) — Primary source: https://huggingface.co/XHToken/Spark-X2.5-4B. A 4B-parameter conversational text-generation model drawing 945 likes and 10.6K downloads on Hugging Face. Ships in Safetensors format using standard Transformers causal-LM architecture. Suitable for local assistant experiments on consumer Mac and Linux hardware.

GitHub Project Radar

DeusData/codebase-memory-mcp — https://github.com/DeusData/codebase-memory-mcp — High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary stars: 42,735; stars_delta_30d: +4,782 (+12.6%) since 2026-08-10; latest_release: v0.10.8 (2026-08-19). Why this is on the radar now: v0.10.8 shipped on 2026-08-19 and the repository was updated on 2026-09-09. Stack improvement angle: Adds a persistent code graph tool surface that MCP-compatible agents (Antigravity, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.
HKUDS/nanobot — https://github.com/HKUDS/nanobot — Ultra-lightweight, open-source, self-hosted personal AI agent framework in Python with WebUI, tools, memory, MCP, multi-agent workflows, automation, and chat apps stars: 47,927; stars_delta_30d: +1,189 (+2.5%) since 2026-08-10; latest_release: v0.3.0 (2026-07-25). Why this is on the radar now: v0.3.0 shipped on 2026-07-25 and the repository was updated on 2026-09-04. Stack improvement angle: Adds a tool surface MCP-compatible agents (Antigravity, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.
ahrm/blender-mcp — https://github.com/ahrm/blender-mcp — An MCP server that exposes Blender's 3D scene controls and Python API directly to connected language models stars: 27,876; stars_delta_30d: +842 (+3.1%) since 2026-08-10; latest_release: v1.1.0 (2026-08-15). Why this is on the radar now: The project was updated on 2026-09-07 with expanded 3D mesh manipulation tools. Stack improvement angle: Adds a creative tool surface that MCP-compatible agents (Antigravity, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.

Extra Research Candidates

IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B — Primary source: https://arxiv.org/abs/2608.03841. Broad open-weights suite including a mixture-of-experts model activating 23B parameters alongside open pre-training corpora.
The Work Now Within Reach: Capability Per Dollar in Production — Primary source: https://openai.com/index/the-work-now-within-reach. An analysis of unit economics showing how declining inference costs unlock high-volume automated business workflows.
OpenAI Fought Dirty on Career-Making Math Problem — Primary source: https://techcrunch.com/2026/09/08/openai-navier-stokes-controversy. The dispute over priority, attribution, and peer verification following OpenAI's Navier-Stokes Millennium Prize proof in Lean.

Show Notes

Episode 112 — September 10, 2026

[00:00] Episode hook

Claude Code CLI shipped version 2.1.267 on npm alongside Hermes Agent v2026.9.7, bringing tighter terminal agent execution and resilient candidate-state gateway management to production developers. Claude Code's update sharpens how the command-line agent handles complex multi-step refactors, delegates subagent tasks, and resolves dirty git diffs without context collisions. Hermes Agent v2026.9.7 complements the front of the release cycle with candidate-state maintenance hardening, isolating plugin and gateway repairs so persistent background sessions stay healthy. Together, these releases point to a clearer operational standard: coding agents must navigate messy local workspaces smoothly while background daemons upgrade without dropping active work.

[02:00] Agent Stack Release Readout: Claude Code CLI 2.1.267; Hermes Agent v2026.9.7

Claude Code CLI 2.1.267 arrived on npm with targeted improvements to terminal-based agentic software engineering. The tool has become a fixture for developers who prefer running autonomous coding agents directly inside their terminal rather than through an IDE wrapper. In this build, Anthropic tightened how Claude Code plans and executes multi-directory file modifications, improving diff hygiene on active git branches and preventing unintended state mutations during test execution. Subagent task delegation has also been optimized to keep prompt contexts compact during extensive codebase exploration. The runtime deployment architecture now manages memory usage more predictably across extended coding sessions, ensuring that long-running tasks avoid token serialization bottlenecks. Terminal feedback mechanisms have also received polish, displaying structured diff views and execution exit codes clearly when complex tool calls complete.

On a parallel release track, Hermes Agent released v2026.9.7 with focused improvements to persistent gateway lifecycle management. Hermes now isolates candidate state during maintenance operations, ensuring that background gateway processes survive configuration and plugin updates without severing connected communication channels. The update also tightens memory reclamation across prolonged agent runs, reducing footprint during heavy tool-use sessions. Configuration options have been expanded to let operators set strict evaluation boundaries, while background watchers report latency and throughput metrics directly to system logs. For engineers running persistent agents in production, these maintenance refinements ensure that long-running automations remain dependable. Both releases emphasize operational resilience: client tools must handle local working tree noise gracefully, while background services must isolate state transitions so active sessions do not fail when configurations change. Together, these enhancements ensure that command-line and persistent development environments remain responsive even during deep recursive directory indexing and extensive test suite execution.

[04:15] Google Antigravity: The Agent-First Platform with Antigravity 2.0, Go CLI, and Python SDK

Google has unified its autonomous developer tooling under the Google Antigravity ecosystem, establishing an agent-first development platform spanning desktop, command line, and programmatic APIs. At the center is Antigravity 2.0, a dedicated desktop application engineered for parallel agent orchestration. Unlike conventional IDE chat panels that treat AI as an inline autocomplete widget, Antigravity 2.0 gives developers a dedicated canvas for coordinating multiple autonomous agents executing tasks in parallel.

The system incorporates an Auxiliary Pane that tracks live background tasks, subagent hierarchies, modified files, and interactive artifacts in real time. For developers in the shell, the Antigravity CLI (`agy`), implemented in Go, delivers near-instant startup times and command execution, making it an ideal harness for scripted automations and terminal-based pair programming. Complementing both is the Google Antigravity Python SDK, which lets teams lease agents programmatically, hook into Model Context Protocol servers, and orchestrate complex multi-agent pipelines with full control over skills and rules. The architecture enforces strict separation between planning and execution, allowing models to formulate multi-step plans before running tool invocations. Observability tooling streams real-time latency and throughput telemetry across subagent networks, providing full auditability for automated code edits.

[06:45] ChatGPT for Mac Rebuilds into Unified Workspace with Computer History and WebMCP Site Tools

OpenAI has delivered a major architectural rebuild of its macOS desktop app, consolidating Chat, Work, and Codex into a unified, native desktop client. The update rethinks how desktop assistants interact with the host operating system. The marquee addition is Computer History, an optional, opt-in feature that constructs a searchable local timeline of user interactions across supported applications and browser tabs. Crucially, OpenAI designed Computer History without capturing screenshots, screen recordings, or audio feeds, relying instead on structured application metadata to give the model rich context about ongoing work without invasive capture mechanisms.

Simultaneously, the desktop app's built-in browser has gained support for WebMCP site tools across Chrome, Brave, Edge, Opera, and Vivaldi. WebMCP allows websites to publish structured action schemas that the assistant can invoke directly, enabling seamless workflows such as inserting comments into shared documents, updating ticket states, or fetching live metrics without manual scraping. Paired with native Google Drive integration in the composer, the rebuilt client transforms ChatGPT on macOS from a conversational sidebar into an active operational workspace. Security controls sandbox browser execution, while local encryption protects stored timeline events from unauthorized access.

[09:10] MiniMax-Text-01 Delivers 4M Context Open-Weights Architecture with Lightning Attention MoE

MiniMax has released MiniMax-Text-01, a 456-billion parameter Mixture-of-Experts architecture activating 45.9 billion parameters per token. Engineered specifically for complex reasoning, autonomous software engineering, and long-horizon agent workflows, the model features a native 4-million-token context window powered by MiniMax's proprietary Lightning Attention mechanism. In traditional Transformer architectures, full quadratic self-attention causes computational complexity and KV cache memory requirements to explode when sequences extend into the millions of tokens. Lightning Attention overcomes this bottleneck by employing chunked linear recurrence, ensuring constant memory footprint during prefill and sustained high token throughput during generation.

The training pipeline for MiniMax-Text-01 combined extensive mathematical reasoning corpora with synthetic execution traces from complex software projects. The resulting checkpoint excels on multi-step reasoning evaluations, coding agent tasks, and long-context needle retrieval across technical documentation. By publishing the weights and enabling flexible inference deployment across high-performance serving frameworks, MiniMax provides developers with an open alternative to proprietary long-context APIs. For teams operating autonomous agents that must ingest entire code repositories, architectural diagrams, and exhaustive log files in a single prompt, the 4M context window eliminates the need for lossy chunking and retrieval pipelines.

[11:35] Mercury 2.5 is a diffusion-style reasoning model, now on OpenRouter

Inception’s Mercury 2.5 is now accessible on OpenRouter, sporting a 260,000-token context window and a 4,096-token output limit. What differentiates Mercury 2.5 from conventional autoregressive architectures is its diffusion-style decoding mechanism. Rather than generating text strictly from left to right, predicting one sequential token at a time, Mercury generates an entire token draft simultaneously and iteratively refines it across multiple passes.

Inception positions Mercury 2.5 as its fastest reasoning model, claiming that parallel token refinement allows the model to solve complex multi-step reasoning problems without leaving accelerator hardware idling on sequential token prediction. The 260K context window accommodates substantial code repositories, technical documentation, and long conversational histories in a single call. As independent benchmarks emerge, developers will be watching closely to see whether diffusion-style parallel refinement matches sequential autoregression on accuracy while preserving its claimed speed advantage.

[13:40] Meta's Muse Agent Wants Your Email, Calendar, and Payments

Meta launched Muse, a personal AI agent designed to act on behalf of users across email, calendars, payment rails, and health services. Rather than functioning solely as a conversational chatbot, Muse is pitched as an agentic assistant that reads personal data, schedules meetings, negotiates appointments, and processes financial transactions within existing personal applications.

The announcement sparked significant developer debate, quickly generating over 500 upvotes on Hacker News. The conversation focused squarely on user consent, data governance, and privacy boundaries. Handing an AI agent blanket access to personal communications, schedules, financial credentials, and health records introduces substantial security risks. Industry observers emphasize that the success of consumer agents like Muse will depend less on raw benchmark performance and far more on granular, per-task permission models that allow users to audit and revoke capabilities transparently.

[15:45] Cognition's $48B valuation says AI coding won't be a one-tool market

Cognition secured a landmark $48 billion valuation in its latest financing round, cementing its position among the top tier of private AI companies. The funding round arrived at an aggressive multiple, surpassing the valuation Cursor commanded prior to its acquisition by SpaceX. The scale of the investment demonstrates that venture capital views AI software engineering as a diverse ecosystem with room for multiple independent players rather than a winner-take-all market.

Software development encompasses vastly different workflows: from interactive, editor-based pair programming to fully autonomous agents executing issue backlogs, generating pull requests, and validating test suites in isolated sandboxes. Cognition's substantial capital reserve allows the company to deepen autonomous capabilities, expand enterprise developer infrastructure, and enhance repository context engines. The continued influx of capital ensures competitive pressure will drive rapid improvements in tooling across the entire developer ecosystem.

[17:35] 1Password says Codex lifted engineering output 21%

1Password reported a 21% increase in engineering productivity following the deployment of OpenAI's Codex across its internal development teams. The case study, published by OpenAI on September 8, details how the password manager's engineering organization integrated the coding agent into daily workflows spanning customer-facing feature development and internal tool maintenance.

What makes the 21% figure notable is 1Password's demanding security posture. As a security-critical software provider, production code must pass stringent static analysis, peer review, and compliance gates. 1Password emphasized that Codex operates strictly within the company's existing security boundaries rather than bypassing established review procedures. The case study provides concrete evidence that autonomous coding assistants can deliver measurable velocity improvements inside organizations where security and code correctness cannot be compromised.

[19:10] Mistral Closes €3 Billion Series D at €21 Billion Valuation

French AI lab Mistral announced the completion of a €3 billion Series D funding round, bringing its post-money valuation to over €21 billion. The announcement captured widespread attention across the global developer community, drawing over 800 upvotes on Hacker News. Mistral has carved out a unique position by championing sovereign, open-weight artificial intelligence, providing downloadable model checkpoints that enterprises and public institutions can self-host and fine-tune.

The substantial capital infusion provides Mistral with the compute resources and research talent necessary to train next-generation frontier models while maintaining its open-weights commitment. In Europe, where data sovereignty and regulatory autonomy under the EU AI Act are paramount, Mistral represents a vital strategic alternative to closed American APIs. The company's forthcoming releases will demonstrate whether sovereign open-weight systems can continue closing the capability gap with closed proprietary frontiers.

[20:55] OpenBMB Ships MiniCPM5-2B, a Sub-3B Open-Weights Model Built to Run On Device

OpenBMB released MiniCPM5-2B, an ultra-compact language model featuring 2.52 billion dense parameters and a native 131,072-token context window. The model posted an average score of 53.9 across 34 standardized benchmarks, outpacing larger competitors including Qwen3.5-4B. MiniCPM5 demonstrates exceptional proficiency in tool use, agentic coding tasks, and needle-in-a-haystack long-context retrieval.

OpenBMB published the complete model weights, intermediate training checkpoints, and datasets under the permissive Apache 2.0 license. To achieve frontier-like capabilities at sub-3B scale, the training pipeline incorporated 400 billion tokens of deep-thinking supervised fine-tuning combined with on-policy distillation from 16 expert teacher models. Quantized GGUF binaries begin at just 1.56 gigabytes, allowing the model to run smoothly on laptops and edge devices via llama.cpp, Ollama, MLX, vLLM, and SGLang.

[22:30] Qualcomm and AWS team up on custom AI inference silicon and 1.6T optical links

Qualcomm Technologies and Amazon Web Services announced a multi-generation partnership to co-design customized AI inference silicon and deploy ultra-high-speed optical networking reaching 1.6 terabits per second across AWS data centers. As AI models scale in parameter size and user request volumes surge, data-center operators are grappling with thermal limits, power constraints, and networking bottlenecks between accelerators.

The collaboration pairs Qualcomm's low-power neural processing architectures with AWS's hyperscale infrastructure. The 1.6T optical links replace traditional copper interconnects, moving massive tensor streams with lower latency and substantially improved energy efficiency across server clusters. By co-designing specialized inference silicon tuned specifically for Amazon's production workloads, the partnership aims to lower the total cost of ownership for high-throughput model serving.

[24:00] PsiQuantum Locks In $100M US Award for Quantum Manufacturing

PsiQuantum finalized a $100 million federal research and development award from the United States Department of Commerce under the CHIPS and Science Act. The award funds the domestic manufacturing and packaging of critical components required for PsiQuantum's fault-tolerant, utility-scale quantum computing systems. The execution of definitive documentation converts an anticipated award into committed federal funding.

Building fault-tolerant quantum computers requires specialized silicon photonic chips, optical control modules, and cryogenic packaging that demand advanced semiconductor manufacturing infrastructure. Anchoring this manufacturing pipeline within the United States strengthens domestic supply chains for emerging deep-tech hardware. As fabrication lines scale, the award positions PsiQuantum to accelerate hardware validation for commercial quantum architectures.

[25:30] OpenAI Posts an AI-Generated Navier–Stokes Proof in Lean

OpenAI published an AI-generated solution to the Navier–Stokes Millennium Prize Problem, coupling an explanatory narrative with a formal, machine-checked proof written in Lean. The Navier–Stokes equations describe the fundamental mechanics of fluid dynamics, such as air turbulence and water flow. The Millennium Prize challenge, which carries a $1 million bounty from the Clay Mathematics Institute, questions whether smooth, well-behaved mathematical solutions always exist in three dimensions or whether singularities can form.

By encoding the formal proof in the Lean interactive theorem prover, OpenAI provided a mechanically verifiable artifact that mathematicians can inspect line by line. While Lean confirms that the logical deductions follow strictly from their stated axioms, the broader mathematical community must still evaluate whether the formal framing accurately satisfies the Millennium Prize specifications and honors prior literature. Regardless of the final mathematical verdict, the milestone illustrates how automated theorem proving is rapidly advancing toward unresolved frontier mathematics.

[27:15] Reducto's r-1 Parses a Whole Page in One Pass at a Penny

Document-parsing startup Reducto launched r-1, an end-to-end multimodal model that parses an entire document page in a single forward pass for exactly one cent. The unified model replaces traditional multi-stage pipelines that chain distinct optical character recognition, layout detection, table extraction, and text formatting models together. Reducto reports a 20% reduction in document parsing errors alongside a significant cost reduction from the previous 3-to-6-cent-per-page range.

In traditional document pipelines, cascading errors across sequential stages frequently corrupt extracted data when layout models misinterpret table boundaries or OCR engines misread noisy scans. By ingesting the entire page image in a single pass, r-1 outputs structured text, tables, layout formatting, and precise bounding coordinates simultaneously. For enterprises ingesting millions of financial filings, legal contracts, and medical records, the predictable one-cent pricing drastically improves the economics of automated document processing.

[29:00] GitHub Project Radar

Codebase-memory-mcp leads this week's developer tooling radar, tracking 42,735 GitHub stars with a 12.6% monthly growth rate. The project compiles a persistent code knowledge graph across 158 programming languages and serves structured queries via a standalone binary over the Model Context Protocol. Connected coding agents can query function definitions, caller hierarchies, and class structures directly instead of repeatedly scanning entire directory trees.

Nanobot follows with 47,927 stars and a fresh v0.3 release. Written in lightweight Python, Nanobot provides an accessible framework for self-hosting personal agents equipped with web interfaces, tool registries, persistent memory, and MCP support. Rounding out the radar is blender-mcp at 27,876 stars, an MCP bridge that exposes Blender’s 3D scene graph and Python modeling commands to connected language models, allowing agents to manipulate 3D geometries and materials programmatically.

[30:45] Model Discovery Check

Inception's Mercury 2.5 is the featured model debut this cycle. Listed on OpenRouter with a 260,000-token context window, Mercury 2.5 employs diffusion-style parallel token decoding rather than conventional autoregressive token generation. By refining full token drafts simultaneously, Inception aims to deliver rapid reasoning throughput on complex mathematical and coding workloads.

[31:30] Local LLM Spotlight

In the local model spotlight, XHToken's Spark-X2.5-4B is garnering strong developer interest on Hugging Face, amassing over 10,000 downloads and 945 likes. Built on standard Transformers causal language model architectures, the 4-billion-parameter model provides an approachable footprint for developers running conversational and lightweight tool-use experiments on consumer workstations.

[32:15] Extra Research Candidates

On the research front, IFM released K2 Horizon, a suite of six open-weight models ranging from 900 million to 375 billion parameters under Apache 2.0, complete with pre-training dataset releases. OpenAI published "The Work Now Within Reach," analyzing how falling capability-per-dollar economics are expanding the frontier of automatable business workflows. Finally, TechCrunch documented the community controversy surrounding OpenAI’s Navier-Stokes Lean proof, where mathematicians are debating priority, attribution norms, and publication etiquette in an era of automated theorem proving.

[33:15] Closing

For complete source links, technical references, and model documentation, visit the show notes at Toby On Fitness Tech dot com. Thanks for listening, and we'll be back soon.

Chapters

00:00 — Intro: Agent Stack Release Readout: Claude Code CLI 2.1.267; Hermes Agent v2026.9.7 / Google Antigravity / ChatGPT for Mac Rebuild
02:00 — Agent Stack Release Readout: Claude Code CLI 2.1.267; Hermes Agent v2026.9.7
04:15 — Google Antigravity: The Agent-First Platform with Antigravity 2.0, Go CLI, and Python SDK
06:45 — ChatGPT for Mac Rebuilds into Unified Workspace with Computer History and WebMCP Site Tools
09:10 — MiniMax-Text-01 Delivers 4M Context Open-Weights Architecture with Lightning Attention MoE
11:35 — Mercury 2.5 is a diffusion-style reasoning model, now on OpenRouter
13:40 — Meta's Muse Agent Wants Your Email, Calendar, and Payments
15:45 — Cognition's $48B valuation says AI coding won't be a one-tool market
17:35 — 1Password says Codex lifted engineering output 21%
19:10 — Mistral Closes €3 Billion Series D at €21 Billion Valuation
20:55 — OpenBMB Ships MiniCPM5-2B, a Sub-3B Open-Weights Model Built to Run On Device
22:30 — Qualcomm and AWS team up on custom AI inference silicon and 1.6T optical links
24:00 — PsiQuantum Locks In $100M US Award for Quantum Manufacturing
25:30 — OpenAI Posts an AI-Generated Navier–Stokes Proof in Lean
27:15 — Reducto's r-1 Parses a Whole Page in One Pass at a Penny
29:00 — GitHub Project Radar
30:45 — Model Discovery Check
31:30 — Local LLM Spotlight
32:15 — Extra Research Candidates
33:15 — Closing

Primary Links

Claude Code CLI npm: https://www.npmjs.com/package/@anthropic-ai/claude-code
Hermes Agent v2026.9.7 release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.9.7
Google Antigravity Documentation: https://antigravity.google/docs
Google Antigravity Python SDK: https://github.com/google-antigravity/antigravity-sdk-python
OpenAI ChatGPT for Mac Changelog: https://learn.chatgpt.com/docs/changelog?type=codex-app
MiniMax-Text-01 on Hugging Face: https://huggingface.co/MiniMaxAI/MiniMax-Text-01
Inception: Mercury 2.5 on OpenRouter: https://openrouter.ai/models/inception/mercury-2.5
Meta: Muse Spark 1.2 on OpenRouter: https://openrouter.ai/models/meta/muse-spark-1.2
Cognition AI $48B valuation: https://cognition.ai/blog
1Password engineering productivity with Codex: https://openai.com/index/1password-codex
Mistral Series D financing announcement: https://mistral.ai/news/series-d
OpenBMB MiniCPM5-2B model release: https://huggingface.co/openbmb/MiniCPM5-2B
Qualcomm and AWS custom inference silicon collaboration: https://www.qualcomm.com/news/releases/2026/09/qualcomm-aws-inference-optical
PsiQuantum CHIPS Act $100M award: https://www.psiquantum.com/news/chips-award-finalized
OpenAI Navier-Stokes Millennium Prize solution in Lean: https://openai.com/index/navier-stokes-lean-proof
Reducto r-1 single-pass document parsing model: https://reducto.ai/blog/r-1
DeusData/codebase-memory-mcp repository: https://github.com/DeusData/codebase-memory-mcp
HKUDS/nanobot repository: https://github.com/HKUDS/nanobot
ahrm/blender-mcp repository: https://github.com/ahrm/blender-mcp
Spark-X2.5-4B local model on Hugging Face: https://huggingface.co/XHToken/Spark-X2.5-4B
K2 Horizon model suite release: https://arxiv.org/abs/2608.03841
The Work Now Within Reach: https://openai.com/index/the-work-now-within-reach
OpenAI Navier-Stokes credit controversy report: https://techcrunch.com/2026/09/08/openai-navier-stokes-controversy

Release Coverage Check

Hermes Agent — Latest stable verified: v2026.9.7, published 2026-09-08T03:12:44Z. Selected missing version(s): v2026.9.7.
Claude Code CLI — Latest stable verified: 2.1.267, published 2026-09-09T18:25:42Z. Selected missing version(s): 2.1.267.
Antigravity CLI — Continuous delivery via @google/antigravity npm package.
OpenAI Codex — Latest stable verified: rust-v0.147.0. Continuous delivery / no new stable release this cycle.

Harness Version Reference

Hermes Agent — v2026.9.7 (stable)
Claude Code CLI — 2.1.267 (stable)
Antigravity CLI — v2.0 (stable)
OpenAI Codex — rust-v0.147.0 (stable)