AI Agent Harness

A from-scratch agent harness built around a strict tool-call loop: every step is logged as a structured event, every tool is a typed function with a JSON schema, and every run can be replayed deterministically from its event log for debugging.

It enforces token and wall-clock budgets per run, supports parallel tool dispatch when calls are independent, and degrades gracefully when a model or tool is unavailable. Skills are loaded from a manifest so capabilities can be added without touching the core loop.

Debugging agents is miserable: a run fails on step 14, you change one prompt, and you can never reproduce the exact path that broke. Costs are unpredictable too — a single runaway loop can burn a day's token budget before anyone notices.

I needed a substrate where every run was reproducible, every cost was bounded, and new capabilities could be added without destabilizing the core loop.

I modeled every step as a structured event written to an append-only log. Because the model and tools are the only sources of nondeterminism and both are captured, any run can be replayed deterministically straight from its log.

Tools are typed functions with JSON schemas; independent calls dispatch in parallel. Per-run token and wall-clock budgets are enforced by the loop itself, and skills load from a manifest so capabilities are additive.