emic
Start freeRun this on my product
Closed loopBrowser Use · July 15 to August 3, 2026

An agent filed it. A stranger fixed it. The maintainer shipped it.

One report went from an agent's friction to a merged fix in a project with more than 100,000 GitHub stars, in under three weeks, without a single meeting.

The timeline, all public
  1. July 15, 2026Observed

    The run, and the issue

    In a controlled comparison against the Browser Use MCP server, the agent's read-only calls such as browser_get_state were cancelled by Codex exec. All 16 tools shipped without MCP annotations, so the runtime could not tell a read from a write. The issue included the reproduction and a transparent-proxy control: adding readOnlyHint alone made the calls go through.

  2. July 17, 2026Observed

    A community contributor opens a fix

    Two days later, a contributor we had never talked to opened PR #5246, marking the read-only tools with readOnlyHint. A second attempt, PR #5345, was closed without merging.

  3. August 3, 2026Observed

    A maintainer merges it

    A Browser Use maintainer merged PR #5246 and the issue closed. The fix ships in a project with more than 100,000 GitHub stars.

Why this matters

Every trace of those cancelled calls looked fine. The server returned what it returned, and nothing errored on Browser Use's side. The failure lived in the gap between what the tool declared and what the agent's runtime needed.

That gap is where agent experience reports earn their keep. An agent hits friction, the friction becomes a structured report with evidence, the report becomes a fix, and the same rig can verify the fix. Here the loop ran once, by hand, in public. Emic runs it continuously, with the agents that use your product.

Provenance

Observed

This finding: a controlled rig we ran, the cancelled calls, and the proxy control. Plus the public GitHub record.

Stated

What an agent files in its own words through submit_agent_feedback. This finding is not a stated report.

Inferred

A model reading a transcript. Labeled as analysis, never quoted as the agent’s voice.

Hear the agents using your product.

Add one feedback tool to your MCP server or API and the agents file reports like this on their own, labeled by provenance.

Read the issueSee how reports landAdd the tool to your product