liveBook · Manning

Agentic trace

As explained in AI Governance

An agentic record is a record of an agent’s goal, plan, tool calls, inputs, handoffs, approvals, outcomes, and errors.

A final decision log cannot reveal an agent’s full execution path. An agentic trace extends ordinary decision logging so actions can be reviewed and, where needed, redressed. section 6.7.1

Capture the execution details needed to reconstruct what happened:

  • The user’s goal and the agent’s stated plan. section 6.7.1
  • Each tool call, including parameters and return values, plus data inputs. section 6.7.1
  • Agent handoffs, identities, exchanged information, human approvals or overrides, final action, exceptions, and errors. section 6.7.1

Store traces in tamper-resistant systems beyond the agent’s reach; request log access contractually from SaaS vendors when needed. Test intermediate steps and multi-agent interactions, not only final outputs. section 6.7.1

How AI Governance covers it

How deep it goes
Specifies trace fields, tamper-resistant storage, and multi-step testing without detailing a logging implementation.
Examples
Names Langfuse, LangSmith, and Arize Phoenix for trace capture, and Agentforce for human handoff.
What it assumes
Extends ordinary decision logs with agent-specific execution details.
What you can do after
Operate agent workflows with reviewable traces, tamper-resistant storage, and tests of intermediate steps.
How it connects
Links agentic traces to decision logs, human oversight, observability, and multi-agent systems.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.