Open-source agent-certify baselines your agent system and alerts on anomalies
CognitiveEngineering · reddit · 2026-10-28
A new open-source tool, agent-certify, learns what your AI agent system normally looks like from past sessions (tools used, communication patterns, error rates, sub-agent behavior) and flags deviations in new runs: unseen tools or first-time sub-agent spawns trigger CRITICAL, while 3x error rates or new agent-to-agent links trigger WARNING.
It also ships a policy engine (mark tools as FORBIDDEN), blast radius analysis for compromised nodes, and an MCP server so agents can check their own anomalies. Works with Claude Code (.jsonl), Cursor (.agent-trace), and MCP logs; installable via pip with 306 tests at 95% coverage.
More from coding & agent
- 90% of a SWE career is complaining about bad codebases — until coding agents arrived — vivekhaldar · 2026-10-28
- Radiology journal publishes review on agentic AI systems, topologies and evaluation — jnkath · 2026-10-28
- Parsers launches: turn URLs from Google Maps, TikTok, G2 into clean JSON via one API — gaganghotra_ · 2026-10-28
- The AI CEO experiment failed: same-model agent teams can't truly diversify — thisguyknowsai · 2026-10-28
- One Grok bot per project: a simple trick to keep agent context scoped — jonathan_wilke · 2026-10-28
- Cursor adds Annotations in Design Mode to direct agents on running apps — testingcatalog · 2026-10-28