Runs & traces

A run is one pass of an agent handling a message — from start until it answers (or fails, or gets cancelled). The Runs page lists every run in your organization; click one to open its trace and see exactly what the agent did.

Runs listCopy Anchor Link

Filter by:

• Agent — narrow to one agent.
• Status — Pending, Running, Waiting for approval, Waiting for input, Succeeded, Failed, Cancelled.
• Time range — a few presets, or a custom range.

Each row shows status, agent, start time, duration, tokens in/out, iteration count, and the related end-user. A run that's running or pending can be cancelled from that row's menu; a run tied to a conversation links straight to it.

TraceCopy Anchor Link

The trace is a step-by-step timeline of a run: every call to the model, every tool call (knowledge_search, MCP tools), and the final answer. When the agent picks a tool that needs confirmation, you'll also see two steps:

• APPROVAL (ask) — the connector's name, the tool's name, the arguments the agent is about to send, and the timeout (5 minutes).
• APPROVAL (decided) — who decided (the end-user, backend API key, or system if it timed out), the outcome (approved, declined, or expired), and whether it was remembered for the conversation.

Similarly, when the agent calls ask_user to ask the user questions, the trace displays QUESTION badges across two steps:

• QUESTION (asked) — the ask_user tool, call ID, expiration time, and the Questions block (header chip, question text, single/multi-select mode, options with descriptions, and the "Other — free text" line).
• QUESTION (answer) — who answered (end-user, API key, or system if timed out), answer timestamp, along with the Answer block (selected options per question chip, or typed free-text response, or the expiration notice if unanswered).

Select a step to see its detail — the prompt sent to the model, the tool result it got back, tokens used at that step.

A running run's trace updates live; you can jump to the latest step while viewing an older one as the run keeps going. A waiting for approval or waiting for input run also updates live — when the user answers or the step expires, the outcome step appears right away.

Trace retentionCopy Anchor Link

A finished run's step-by-step content — the prompts sent to the model, and its tool calls and their results — is removed once it's older than the number of days set under Usage & quotas — 30 days by default, set for your organization by the Harnix operator. Deletion runs once a day.