Building reproducible LLM agents with strict determinism guarantees

I’ve been building a deterministic agent system in TypeScript with the explicit goal of making agent behavior auditable, reproducible and bounded — something that most agent frameworks currently ignore or treat as optional.

Key properties: • Bounded tool loops with fixpoint convergence or max_iterations • Deterministic planners (det-tools, mock, llm-live stub/real-path) • Reproducibility via plan/execution/finalTrace hashes + replay bundle v2 • Surfaced violations instead of silent failures • Capability-driven tool selection (latest addition) • HTTP API (/agent/run, /tools, /capabilities) with negative test coverage • One-command demos (npm run demo:agent:llm-live:real, etc.)

It’s still early (v0.2.0, 165 commits in ~12 days), no LLM live full yet, no npm package, but the bounded/replay/determinism core is already verifiable via tests and demos.

Repo: https://github.com/crasofuentes-hub/deterministic-agent-syst...

Would love feedback from people working on reproducible agents, AI safety evals, or enterprise workflows where non-determinism is a real problem.

Thanks!