AgentInspect: open-source TypeScript tool to record and verify agent execution paths locally
Evening-Let-7562 · reddit · 2026-10-26
The author released AgentInspect, an MIT-licensed TypeScript library and CLI for recording agent runs locally and checking their execution paths.
- Core demo: two runs return the same refund-policy answer, but one generates the answer before retrieving the policy — final-text comparison passes both, while an explicit ordering rule rejects the reversed run.
- Workflow: local JSONL → execution tree → named check failures → an evidence bundle you can review before attaching to an issue or PR.
- Design choices: scripted responses mean no API key or cost; harmless path changes (e.g., repeated retrieval in allowed order) still pass.
- Limits: it checks recorded events and rules you supply; calling the right tool doesn't prove the answer used its result. LangSmith and Promptfoo already support trajectory checks — this differentiates on local debugging and evidence workflow.
Repo: github.com/rajudandigam/agent-inspect
More from coding & agent
- Celesto pitches a one-line GitHub Actions swap claiming ~90% cheaper CI minutes at agent scale — aniketmaurya · 2026-10-26
- After 10,000+ hi-fi UI mockups, a 10-year designer shares 8 taste principles and 5 AI design skills — xiaohu · 2026-10-26
- Dev builds 'Physical AI OS' in a weekend where agents write and control robots — chrismatthieu · 2026-10-26
- SoundBit: open-source chiptune workstation with a built-in MCP server for live AI composing — RenasDev · 2026-10-26
- 50-year-old dad dives into AI agents, starting with HermesAgentDesktop — Teknium · 2026-10-26
- Matt Shumer launches email-based AI employee for SMBs with its own browser — mattshumer_ · 2026-10-26