AgentDesk

AgentDesk gives teams a single place to deploy, monitor, and audit AI agents. It ships with a replay viewer, token-cost dashboards, and guardrail presets so non-ML engineers can ship agents safely.

Teams adopting agents had no visibility into what prompts ran, what they cost, or where they went wrong. Debugging meant reading raw logs.

I designed an event-sourced execution log that replays any agent run step by step, plus a cost tracker that attributes spend per task.

A FastAPI orchestrator emits events to Redis Streams; a React dashboard subscribes for live updates. Guardrails are expressed as composable middleware.

• Capturing enough context for replay without exploding storage
• Designing guardrails that are strict yet not paralyzing
• Supporting both cloud and fully air-gapped deployments

Currently in private beta with three design partners. Early signal shows a 60% reduction in time-to-debug agent failures.