Deep architectural teardowns of AI agent frameworks — how they actually work under the hood.
These aren't surface-level reviews or marketing summaries. Each teardown is based on direct source code analysis, cross-referenced against documentation, with every claim backed by specific file paths and function names.
| Agent | Latest Version | Last Reviewed | Source Repo |
|---|---|---|---|
| OpenClaw | v2 (Apr 2026) | 2026-04-01 | openclaw/openclaw |
| Hermes Agent | v1 (Apr 2026) | 2026-04-01 | NousResearch/hermes-agent |
| Comparison | Date |
|---|---|
| OpenClaw vs Hermes Agent | Apr 2026 |
Every teardown follows the BLUEPRINT.md — a systematic framework for analyzing AI agent architectures. It covers:
- Tier 1 (Critical): Agent loop, prompt assembly, memory architecture, context management, security, tool system
- Tier 2 (High): Sessions, skills, channels, provider management
- Tier 3 (Medium): Sub-agents, scheduling, sandboxing, extensibility
- Tier 4 (Esoteric): Node/device systems, research infrastructure, editor integrations
Every section is verified against actual source code, not just documentation. Discrepancies between docs and code are flagged explicitly.
Use the template to start a new teardown. It follows the blueprint structure with placeholder sections.
- Not a "which agent is best" ranking
- Not a getting-started guide for any framework
- Not affiliated with any of the projects reviewed
Found something wrong or outdated? Open an issue or PR. Teardowns are living documents — agent frameworks evolve fast.
Roy Osherove — exploring how AI agent systems work at the architecture level.
All teardowns are based on publicly available open-source code.