Use agent run traces to diagnose behavior without replaying actions

Start with a concrete diagnostic question

A trace is most useful when you ask what the agent actually observed, which tools were available, what action it proposed and what result it received. Lupin records correlated diagnostic activity for new agent turns across supported channels. The console’s Developers → Agent activity view shows recent traces; earlier conversations are not backfilled. Do not treat a trace as an instruction to replay the action. Its purpose is to explain observed behavior while the durable approval and operation records remain the authority for authorization and external outcomes.

Follow the identities across stages

A turn has a trace identity and associated conversation and turn identities derived from authenticated context. Model events can show selected history, available schemas, outputs and reported token usage; tool and API events link execution intervals and typed results. These IDs help distinguish a repeated handling attempt from a new business action. Do not replace server-derived workspace binding with a model-selected organization ID. Repeated ingress can retain the same trace identity while recording distinct handling attempts, which is useful when diagnosing retries without pretending each attempt was a new user request.

Know what content capture means

Default storage is metadata with three-day retention. Owners and administrators can choose metadata or bounded redacted content and one to seven days in the dashboard. Redacted capture excludes hidden reasoning, audio and raw provider exception detail, but arbitrary secrets cannot be recognized perfectly from natural language. Metadata mode removes diagnostic content at ingestion before persistence; it does not stop the agent’s normal message processing. Enable content capture only for a concrete diagnostic need, keep it private and understand that switching back affects future writes while existing captured content expires under its policy.

Read timings and missing evidence honestly

Root, model, tool and API intervals can overlap, so adding their durations can produce an invented total. Some provider timings are reported rather than directly measured, and Worker clocks can differ. A missing audio or handset-delivery measurement should remain not measured, not zero. Background diagnostic export can be lost because of timeout, termination or storage failure, so an absent trace does not prove no action happened. An unfinished root trace also does not become successful merely because no explicit failure event arrived.

Keep access and accounting distinct

Private conversation traces require an authorized unscoped owner-capable key at the API boundary, and console access rechecks the signed-in human’s owner or administrator membership. Project and inbox keys do not gain trace access through ordinary resource grants. Diagnostic retention is separate from email content retention and durable usage accounting. Traces are operational telemetry, not a cryptographically signed audit ledger, billing invoice or recovery interface. No diagnostic write failure should trigger a second email, call or tool execution.

Investigate before changing the prompt

For a comprehension issue, compare the actual selected context, instructions, schemas, model output, tool result and final delivered response. A missing source may explain the behavior better than a model setting. Your next step is to inspect one recent trace with a narrow question and record the evidence gap, avoiding any retry or action button because diagnostics cannot authorize one.

Sources & verification

Reviewed against checked-in Lupin API handlers and documentation. Example requests were not executed against live providers.

• Lupin authentication and owner boundaries ↗Checked 2026-11-05
• Lupin MCP results and approvals ↗Checked 2026-11-05
• Lupin durable usage accounting ↗Checked 2026-11-05

Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.