Practitioner: Tool-Call Trajectories Are Higher-Signal Than CoT for Agent Monitoring
joshua_saxe · x · 2026-09-04
Drawing on hands-on experience, Joshua Saxe argues CoT monitoring is useful but lower-signal than simply looking at the n-gram of tool calls up to time t and gating the next tool call on what the trajectory suggests.
A concrete alternative from a security practitioner: monitor observable behavior sequences instead of the model's inner monologue for agent anomaly detection. He tags another researcher for input.
Related event: Security veterans untangle AI safety: security vs alignment vs policy(9 posts)→
More from coding & agent
- How do you evaluate GenAI output quality before shipping? Devs weigh in — ssnuopi · 2026-10-21
- RunInfra launches $10/mo coding plan: fastest open models across 19 agents incl. Claude Code — ycombinator · 2026-10-21
- Matt Shumer launches an 'AI employee' you assign work to over email — mattshumer_ · 2026-10-21
- uv-ty integration: how to enable PEP 723 script type checking — charliermarsh · 2026-10-21
- Astral ships experimental uv-ty integration: uv check type-checks your project — charliermarsh · 2026-10-21
- Dev's self-built tracker claims 7.1x speedup with Claude, 2.5 work-years banked — arthurcolle · 2026-10-21