Reasoning trace

The recorded sequence of an agent's intermediate reasoning steps from initial prompt to final action, including the system prompt, retrieved context with provenance, model output (including any chain-of-thought the platform exposes), planned action, tool calls with parameters and responses, and the executed action. The reasoning trace is the agent equivalent of a stack trace and is the primary forensic artefact when an agent action is queried by a regulator, an auditor, or an incident responder.

Reasoning trace is the field most commonly missing from vendor-native agent logging in 2026. Microsoft, Anthropic, OpenAI, and Google all capture deployment ID, agent identity, and final tool-call chain natively; the intermediate-conclusion and planned-vs-executed distinctions almost always require deployment-layer instrumentation. Enterprises relying on vendor-native logging discover the gap during the first regulator inquiry under EU AI Act Article 73, when the inquiry asks 'why did the agent take action X' and the answer requires inference from incomplete records.

Related frameworks

• MTTD-for-Agents →

Articles that analyse this term

• Agent incident response: the six-step playbook for when an autonomous-AI deployment breaks production
• Agent observability stack: the four layers production agentic-AI actually needs (and what each one misses)
• EU AI Act Article 12 audit-evidence template for agentic AI

Primary sources

• European Union. EU AI Act Article 12 — Logs
• Anthropic. Claude API — message tracing