Reference implementation separates model judgment from system authority in AI agents
Dapper-Roof2370 · reddit · 2026-09-17
A developer published a small FastAPI/Pydantic reference project addressing a common production agent failure mode: mixing model judgment with system authority.
- Flow: task → typed validation → worker → policy boundary
- Policy layer routes to ALLOW, HUMANAPPROVAL, HUMANREVIEW, or REWORK
- High-risk actions require approval; low-confidence results get review; missing evidence triggers rework
- Deterministic worker runs without an API key; automated tests cover the routing logic
- Author asks operators what to add next: durable state, idempotency, scoped identity, or audit logs
More from coding & agent
- Quanta: Open-Source Python Library Unifies Quantized Vectors, BM25 and Graph RAG — _reachsumit · 2026-09-17
- Thread Argues C, Not Python or Rust, May Be the Key Language of the AI Era — solyarisoftware · 2026-09-17
- RAG paper finds stronger retrieval beats bigger generators, proposes DRAG query-adaptive routing — _reachsumit · 2026-09-17
- DeepSeek V4.1 Flash gotcha: Pi agents need explicit "input": ["text", "image"] config — solyarisoftware · 2026-09-17
- Open-source Skill Index compares MCP config drift across Claude, Codex and Cursor — CombinationOk2374 · 2026-09-17
- One-command MCP connector wires new Jev model into Claude Code and Codex — purealgo · 2026-09-17