I'm Emmanuel and I built Hiver.
I've been building agents for a while, and I've found that the hard part isn't getting them to perform actions. It's understanding why they fail.
A missing tool, a misconfigured permission, an ineffective skill, or a bad system prompt can cause failures several steps later. By that point, ordinary logs often don't provide enough context to identify the original problem.
I started thinking about what a developer tool for this should look like. My idea was that the runtime itself should produce a complete trace of the agent's execution, and that the trace should be replayable and inspectable. Eventually, I'd also like coding agents to use these traces to diagnose and fix their own failures.
Hiver runs agents in isolated environments and records their terminal activity, browser state, file operations, network requests, tool calls, and approvals as one timeline.
On hiver.sh, click 'play' to open an interactive replay of a real agent run. This agent uses a browser to complete 2FA, checks email, generates a report, finds an available time, schedules a meeting, and replies with the report attached.
The workflow uses two sandboxes and has a lot of moving parts, which is exactly what I wanted to demo.
Do these problems sound familiar? How are you tracing and debugging agent failures today?