Agno v3.1.12
Agno v3.1.12 introduces native approval flows for Claude and Codex harness agents, initially disabled company routines, chart headline values, and fixes for FastAPI tracing, routine cancellations, and knowledge chunking.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
Agno release v3.1.12 brings several architectural enhancements and bug fixes across AgentOS, agent harnesses, and orchestration tools: - Native Approvals for Harness Agents : Claude and Codex permission requests now interface directly with AgentOS approval UI. When backed by a database, native requests persist a PAUSED run state visible in Approvals and Chat views, resuming execution via 'Resolve plus Continue Run' without replaying prompts. Note: CodexAgent without approval settings now defaults to the native on-request policy and pauses instead of failing silently, governed by hitl timeout (default 900s). - Initially Disabled Company Routines : Added support for Routine(enabled=False) , allowing templates to initialize inactive routines that persist across restarts, remain visible in listings, and are ignored by the scheduler until activated. - Chart Headline Numbers : ChartTools.create chart() now accepts an optional formatted value string rendered alongside chart titles in Agno OS. - FastAPI Native Tracing Fix : Fixed an issue where FastAPI 0.142.0+ OpenTelemetry traces polluted Agno run views when AgentOS(tracing=True) was enabled. The application now opts out of FastAPI request spans and filters foreign spans to ensure traces root at the Agno run. - Company Routine Cancellation : Corrected premature cancelled state emission in POST /company/routines/{id}/cancel to wait for actual durable executor settlement before updating run status. - AccuracyEval Updates : Replaced default judge model from the retired o4-mini to gpt-5.6-terra and transitioned related accuracy…
可执行摘要
Agno v3.1.12 adds native approval persistence for Claude and Codex harness agents via AgentOS UI, changing default CodexAgent policy to 'on-request'. It adds Routine(enabled=False) , fixes FastAPI tracing contamination under tracing=True , patches Knowledge Base per-upload chunker state pollution, and updates default AccuracyEval judge models to gpt-5.6-terra using OpenAIResponses .
- Agent 实用度
- 62/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Audit any CodexAgent instances running without explicit approval settings, as they now default to 'on-request' and will pause runs awaiting human approval instead of dropping requests.
- Verify FastAPI trace configurations if running AgentOS with tracing=True to confirm foreign request traces no longer clutter the Agno Runs view.
- Update AccuracyEval configurations or downstream tests relying on the default judge model to handle the new default model gpt-5.6-terra and OpenAIResponses wrapper.