PydanticAI 2.57.0 adds ApprovalReviewer: a second model pre-clears safe agent tool calls
KhuyenTran16 · x · 2026-11-12
Khuyen Tran covers ApprovalReviewer, new in PydanticAI 2.57.0, which tackles human-approval fatigue in agents.
Problem: human approval is a good safety check for dangerous shell commands, but when every harmless command also needs approval, repeated prompts train people to blindly click 'Approve'.
How it works: a second model (the example uses anthropic:claude-haiku-4-5) reviews each approval-required tool call before a human sees it, with a customizable question like 'could this delete data, reach the network, or expose secrets?'
Logic: low-risk calls continue in the same run step; risky, uncertain, or failed reviews still require human approval.
More from coding & agent
- Creator shares Claude-powered video workflow: the model keeps learning and it compounds — floguo · 2026-11-12
- LangChain releases fireside chat with Modal CEO on sandboxes and inference at Interrupt NYC — LangChain · 2026-11-12
- With just a webcam and minutes of calibration, Fable 5.2 builds a tool that tracks what you read — ChrisGPotts · 2026-11-12
- Profound CTO on Fractal Speed: a marketing harness cut two weeks to two hours — soleio · 2026-11-12
- 44 agent skills distilled from shipping 3 live products, open-sourced on GitHub — PawelHuryn · 2026-11-12
- GitHub's free Product Engineering skills walk you from idea to a deployed agentic product in 2-3 weeks — PawelHuryn · 2026-11-12