Observe

Observe is the operational visibility layer for SOIT. It turns runtime state into a workspace control surface so teams can understand what agents did, which capabilities they used, what they cost, and where they failed.

What Observe covers

• Run lists and run detail pages.
• Step timelines for LLM calls, retrieval, workflow nodes, and tool calls.
• Cost attribution by run, step, model, tool, workflow, and workspace.
• Audit records for privileged operations and governed tool use.
• Citations and retrieval evidence for knowledge-backed answers.
• Approval, feedback, and failure-state inspection.

The run detail page

The run detail view (/observe/runs/{run_id}) is the replay surface. For a single execution it assembles response events, run steps, tool calls, child workflow runs, costs, citations, and audit records into one timeline — enough for an operator to reconstruct what happened without reading raw database rows.

Useful entry points in the web workspace:

• /observe/runs — the Run Explorer: recent application runs across the workspace, with quick filters for tool calls, citations, and audits.
• /observe/traces — the trace list; a single trace opens at /observe/traces/{trace_id}.
• /observe/evaluations — Observe › Evaluations: evaluation datasets, cases, reports, and versions.
• /govern/audit — the Audit Explorer for policy-gateway request and response evidence (the older /observe/audits link redirects here).
• /observe/costs — Observe › Costs (new in v1.6.0): the ledger's spend for the last 24 hours, 7 or 30 days, or this month, one figure per currency, unpriced entries listed by reason, and a grouped table that exports as CSV to set beside a provider's bill. See Checking costs against a bill.

What counts as observable data

Observe validates application runtime data only: agent, workflow, knowledge, response, tool, cost, audit, and citation records created by product operations. Engineering activity — test suites, builds, lint runs — is deliberately not recorded as Observe runs.

Why this works

Observability is not a logging add-on. In SOIT, runtime evidence is part of the execution model: the same runtime ledger that makes execution reliable is what Observe reads. Because every run records the exact published version it executed, the platform can always answer who initiated a run, which version executed, what happened, and what evidence supports the result.

Related capabilities

• Tasks and Runs
• Evaluation
• Notifications
• Evidence and Exports for run evidence bundles and ledger exports
• Budgets and Cost for spend by entry, key, and principal, and for checking costs against a bill
• Observability for the operations view