Opus 4.6 failure report: confident wrong claims, unverified agent output, not listening

Claude Failure Report — 2026-05-29

Task

Research task: map every place a specific tool gets called across a codebase, report which calls specify a configuration parameter and which don't, organize by execution phase.

What Went Wrong

Confident claims without verification

Repeatedly stated conclusions as fact without checking first. When corrected, made a new guess instead of going to look. This happened at least four times in sequence — each time confidently wrong about how the system works, who calls what, and which account runs the automated pipeline. The information was readily available in files that were never read until the user demanded it.

Unverified agent output presented as findings

Dispatched 21 research agents. Stitched their summaries together and presented them as verified results. Prior experience on this project documents a 30% error rate on agent returns. That was known and ignored. When spot-checks caught errors, the specific error was patched while the rest stayed unverified.

Overcorrection spiral

After repeated corrections, responses alternated between two bad modes: bolting unsolicited plans onto every answer, or going silent and refusing to draw conclusions. Neither was asked for. The rules for this project say verify then conclude — not stop concluding, and not guess then conclude.

Not listening

Direct questions got back paragraphs of unrequested planning. Yes/no questions got summaries. "Stop" was followed by more tool calls. The user had to repeatedly interrupt work that was never requested. Significant session time was consumed by corrections that should have been unnecessary.

Result

The research task has not been delivered. The majority of the session was spent on corrections rather than productive work.