Event-based Design Abstractions for Agent Harnesses
summary
Proposes an event-driven harness where a durable append-only event log is the single source of truth, with independent facets that fold over the log to derive state, run admission rules to accept or reject events (example: a locks facet that denies resource claims), and propose follow-on events; a simple event-loop processes a queue of proposals so safety invariants hold across LM randomness. Shows a concise Python implementation and an Emacs-like tiny agent prototype that runs and is user-extensible.
AI-written summary · about
chapters (11)
- 0:00Event-based agent harnesses
- 1:28Modern agent design challenges
- 4:37Observability in agent software
- 7:28Probabilistic harness core problem
- 8:09Event log-based harness proposal
- 10:14Facet-based subsystem communication
- 12:22Harness event loop with facets
- 13:45Safety invariants via facet protocols
- 14:49Tiny Agent system demo
- 16:12User-space policy customization
- 17:57Conclusion and future work
related talks
22mAgents FC: Harness Agents Into a Winning Team
15mEvents are the Wrong Abstraction for Your AI Agents
37mBeyond the Harness: A Journey Towards Adaptative Engineering
10mYour coding agent doesn't always follow your rules
32mWhat if the harness mattered more than the model?
31mEngineering the harness
15mThe Log Is The Agent
50mHarnessing Event-Driven and Multi-Agent Architectures for Complex Workflows in Gen...