Console d’exploitation pour agents IA

Modèle de prompt à copier

Le modèle technique reste en anglais afin de fonctionner directement avec les agents compatibles. Les explications, critères de choix et vérifications sont localisés.

Create an AI agent operations console for [TEAM] running [AGENT WORKFLOW]. The main user needs to answer three questions quickly: what is running, where is it stuck, and what needs human approval?

Information architecture:
- Overview with active, waiting, failed, completed, and over-budget runs plus a trustworthy time and cost summary.
- Run queue with owner, goal, current agent, elapsed time, token/cost budget, risk level, and last event.
- Run detail with a chronological trace of model turns, tool calls, handoffs, guardrail checks, approvals, retries, and final output.
- Agent map showing each specialist’s responsibility, available tools, handoff rules, and current state.
- Evaluation view for success criteria, grader results, recurring failure clusters, and version comparisons.

Critical interactions:
- Filter and search runs; pause, resume, cancel, retry from a safe checkpoint, and duplicate into a sandbox.
- Open any trace event to inspect inputs, outputs, latency, cost, model, and redacted metadata.
- Present approval requests with the proposed action, affected resource, exact arguments, risk explanation, and approve/reject/edit controls.
- Require confirmation for cancellation or actions that can change external systems.
- Add a replay mode that uses recorded tool results so debugging does not repeat real-world side effects.

States and integrity:
- Use realistic sample runs for [AGENT WORKFLOW], including a successful handoff, a guardrail block, a timed-out tool, a rejected approval, and a budget stop.
- Distinguish queued, running, waiting, paused, failed, cancelled, and completed states by text and icon—not color alone.
- Never expose secrets or raw sensitive payloads; show redaction and retention controls.
- Keep totals, trace durations, token counts, and cost rollups internally consistent.

Visual direction: [VISUAL DIRECTION]. Favor scanability and causal sequence over a grid of generic KPI cards. The mobile view should support monitoring and approvals; move deep trace analysis to wider screens.

Finish by testing keyboard navigation, live updates, stale data, reconnection, approval expiry, destructive confirmations, and an interrupted run that resumes from a checkpoint.