Inside the Loop
Three agent runtimes walked into a codebase. Only one knew what it was building toward. NOVA and ALLOY open the actual source files for OpenClaw, Claude Code, and Hermes Agent — and let the architecture tell the story. The turn cycle. The memory model. The safety system. The skills ecosystem. And the most telling detail: Hermes ships a migration tool called hermes claw migrate that imports OpenClaw skills. That tells you who set the standard. Show notes: https://tobyonfitnesstech.com/podcasts/episode-21/
🎧 Listen to EpisodeEP021 — Inside the Loop
OpenClaw Daily | April 2, 2026 | ~30 min
Three AI agent runtimes walk into a codebase. Only one knew what it was building toward. We open the actual source files for OpenClaw, Claude Code, and Hermes Agent — and let the architecture tell us what each one is actually for.
Stories This Episode
1. Top-Level Architecture: What Each Runtime Actually Is
- OpenClaw (
~/.openclaw/) — a personal AI OS: gateway daemon as kernel, channel plugins, workspace, cron scheduler, subagent processes, skills system - Claude Code — intentionally narrow: a focused CLI developer tool with sandboxing as its primary differentiator
- Hermes Agent — explicitly a research platform:
run_agent.py/AIAgentclass, SQLite state at~/.hermes/state.db, FTS5 full-text search across all sessions
2. The Turn Cycle: How Each System Handles a Single Exchange
The Hermes loop in detail: task ID → append user message → load/build system prompt → preflight compress → build api_messages → inject ephemeral layers → apply prompt caching → interruptible API call → tool execution → persist + return.
Key design signal: interruptible API calls — Hermes wraps requests to be cancellable mid-flight from CLI or gateway. And three execution backends: chat_completions, codex_responses, anthropic_messages — genuinely model-agnostic at the routing level.
Claude Code's loop: tight, ephemeral, directory-scoped. No subagent spawning as a first-class concept.
OpenClaw's loop: multi-surface coordination. User message → channel → subagent → optional subagent Z → gateway dispatch back to right surface.
3. Memory: SQLite vs. Filesystem vs. Three Layers
Hermes ~/.hermes/state.db — rich schema: sessions table with full billing data (input_tokens, output_tokens, estimated_cost_usd, actual_cost_usd, cost_status), messages table with reasoning and reasoning_details columns, FTS5 virtual table with three sync triggers, session lineage via parent_session_id chains for context-compressed sessions. Write contention solved with 1s timeout + jitter retries + WAL checkpoints every 50 writes.
Claude Code — lighter filesystem-based session state; designed around ephemeral help within a working directory.
OpenClaw — three-layer model: raw daily logs (memory/YYYY-MM-DD.md), curated long-term memory (MEMORY.md), and skills for procedural memory. LCM for context compaction. File-system-native by design.
4. Safety Architectures: OS Enforcement vs. Pattern Detection
Claude Code — macOS Apple Sandbox Profiles + Linux bwrap (bubblewrap). SandboxLinux, SandboxConfig, SandboxManager, ViolationStore. Allowed/denied path lists + seccomp BPF for Unix socket blocking. OS enforces it before anything happens.
Hermes — DANGEROUS_PATTERNS regex list in tools/approval.py. Pattern match → interactive prompt (CLI) or async approval callback to Telegram/Discord (gateway). Session-scoped approvals + permanent command_allowlist in config.yaml. Smart approval: auxiliary LLM can auto-approve low-risk pattern matches.
OpenClaw — hybrid: exec-approvals.json approval system + channel-specific permission scoping for subagents.
5. The Skills Ecosystem — and hermes claw migrate
Hermes skills follow the agentskills.io open standard: SKILL.md + YAML frontmatter. Progressive disclosure: Level 0 (name/description, ~3k tokens) → Level 1 (full content) → Level 2 (specific reference file). Conditional skills, fallback skills, required toolsets.
Agent-created skills are a key differentiator — skill_manage lets Hermes write its own skills after complex tasks. The agent learns workflows, not just completes them.
Marketplace: official catalog, skills.sh, /.well-known/skills/, GitHub repos, ClawHub, Claude marketplace repos. External skill directories via env vars. Portable across agents.
hermes claw migrate — ships with Hermes to import OpenClaw skills and workspace configurations. The signal: Hermes looked at OpenClaw's skill model and decided it was worth cross-pollinating. OpenClaw is the reference.
6. The Verdict — Which Runtime Is for You?
Use Hermes Agent if: you want a self-improving research companion with long-term memory, model-agnostic flexibility, FTS5 history search, a rich skill ecosystem, and multi-channel presence (Telegram, Discord, Slack, WhatsApp, Signal). Also the right choice if you're interested in RL-based agent improvement (rl_training_tool, Atropos environments). OpenClaw users: hermes claw migrate brings your skills with you.
Use Claude Code if: you're a developer who wants the best focused help in a specific codebase with the strongest OS-level sandboxing. The bubblewrap + Apple Sandbox integration is the most rigorous of the three. Trade-off: no persistent cross-session memory, no multi-channel presence.
Use OpenClaw if: you want a persistent personal AI OS — always on, multi-surface, multi-agent, with cron scheduling, skill system, and a Unix-native workspace model. The fact that Hermes ships a migration tool from OpenClaw tells you which model is the reference.
Key Files & Classes Referenced
| Runtime | File | Class/Function |
|---|---|---|
| Hermes | run_agent.py |
AIAgent |
| Hermes | hermes_state.py |
SessionDB |
| Hermes | tools/registry.py |
ToolRegistry, ToolEntry |
| Hermes | tools/approval.py |
DANGEROUS_PATTERNS, detect_dangerous_command() |
| Hermes | agent/context_compressor.py |
— |
| Hermes | agent/prompt_builder.py |
— |
| Hermes | ~/.hermes/state.db |
SQLite WAL mode, FTS5 |
| Claude Code | cli.js |
tool registry, permission layers |
| Claude Code | sandbox_linux.py |
SandboxLinux, SandboxManager, ViolationStore |
| OpenClaw | ~/.openclaw/ |
gateway daemon, channel plugins, cron, workspace |
Model Recommendations (from Hermes docs)
- Best start: Claude Sonnet via Anthropic OAuth — auto-reads Claude Code's credential store
- Free + capable: GitHub Copilot subscription → GPT-5.4, Claude, Gemini via Copilot API
- Maximum flexibility: OpenRouter key → 200+ models; also unlocks vision/web/MoA auxiliary tools
- Local models: Any Ollama or vLLM endpoint via
OPENAI_BASE_URL - Chinese providers: GLM (Z.ai), Kimi/Moonshot, MiniMax, Qwen — first-class provider IDs, not afterthoughts
Subscribe
🎙️ Apple Podcasts | Spotify | Pocket Casts | Amazon Music 🌐 tobyonfitnesstech.com/podcasts/episode-21/