PydanticAI v2.53.0 adds ToolCallJudge to vet risky agent tool calls before execution
KhuyenTran16 · x · 2026-10-15
PydanticAI v2.53.0 introduces ToolCallJudge: agents connected to external tools can now have risky tool calls (deleting files, leaking sensitive data, other irreversible actions) reviewed and blocked before execution, adding a safety gate ahead of side effects.
The release also patches a high-severity bug (GHSA-6fqq-452j-qhrp): streamed requests through ConcurrencyLimitedModel / limitmodelconcurrency could keep their concurrency slot after early exit or cancellation, letting repeated streams block every request sharing the limiter. The fix changes limiter sharing semantics—model wrappers now raise UserError when sharing a limiter with the requesting agent, acquire() takes a slot on every call, and custom limiters must allow cross-task release(). v1 is unaffected.
Related event: PydanticAI 2.53.0 Adds ToolCallJudge to Screen Risky Tool Calls(2 posts)→
More from coding & agent
- vLLM's bitwise-consistent RL: zero KL on Nemotron NVFP4, 91.3% throughput on DeepSeek-V4-Flash — DynamicWebPaige · 2026-10-15
- vLLM achieves bitwise-consistent RL training: DeepSeek-V4-Flash hits 0 diff over 100 steps — vllm_project · 2026-10-15
- ChatGPT Sites + MCP Events + dots: A Workflow-Triggered Agent Pattern — gabrielchua · 2026-10-15
- Dev reflects on hand-refactoring with Opus 4.6 instead of trusting agents — madeofAjala · 2026-10-15
- Why games are a real, taste-aware coding benchmark, not just a gimmick — vitaliychiley · 2026-10-15
- OpenUI 1.0 brings ChatGPT-style Intelligent UI to local open models, MIT-licensed — dr_cintas · 2026-10-15