Skip to content
Agent Forge Docs

Observability: Trace and Probe

Observation splits a run into queryable spans: model calls, tools, agents, approvals, and more. Write failures do not block the main path by default, so telemetry cannot take down a session; list and detail APIs serve the console and troubleshooting.

Probe

Section titled “Probe”

Probe.start / Probe.end record one observation. Types include generation, tool, agent, approval, span, event, and others. Entry source can distinguish chat / task / command / shell / trigger / summarize, and so on.

flowchart LR
  Run[Session / trigger run] --> G[generation]
  Run --> T[tool]
  Run --> A[agent]
  Run --> P[approval]
  G --> Store[(observation)]
  T --> Store
  A --> Store
  P --> Store

Query

Section titled “Query”
  • GET /observation/traces: filter by session, agent, team_run_id, time, status; detail includes flat list and tree
  • Additional cost and approval aggregations, plus an Agent Graph view

Observation is not the audit hash chain: the former answers “what happened and how long”; the latter answers “were records rewritten.” Accountability still rests on hash-chain audit.

Related reading: