Coding agents are getting good at producing code. That is no longer the hard part.

The hard part is knowing whether the agent understood the task, stayed inside the right scope, changed the right files, avoided unnecessary complexity, and produced work that can be trusted before it reaches a pull request.
This talk introduces task contracts as a practical control layer for AI coding agents.

A task contract defines what the agent is allowed to do before it starts: intent, allowed scope, forbidden areas, expected evidence, success criteria, required checks, and repair instructions.

I will walk through failure modes I have seen while building Cambone, a local-first AI coding quality gate: no-op changes, scope drift, overengineering, false completion, weak attribution, and agents passing benchmarks while still failing real repo context.

The audience will leave with a concrete task-contract structure they can apply to their own coding-agent workflows.
The core argument is simple: AI coding does not need more vibes. It needs contracts, checks, and control.