The Handoff Is the Work

GitHub put GPT-6 Astra into Copilot and gave it the language every software leader wants to hear: plans, validates, confirms.

Good. That is useful.

It is also where the easy mistake starts.

When a coding agent can check its own work, teams will be tempted to treat review as less important. The agent ran the tests. The agent fixed the lint. The agent says the task is done. Move faster.

Maybe. But only if the handoff got better too.

Self-checking is not the same as self-governance. An agent can validate against the instructions, tests, and tools it was given. It cannot know which production scar changed how your team treats billing, permissions, support flows, customer data, migrations, or anything else with real blast radius.

That knowledge lives in the organization.

If the agent says it is done, the next question is not whether the code looks clever. The question is whether the claim is reviewable. What changed? Why? What did it test? What did it skip? What assumptions did it make? What boundary did it touch?

A clean handoff lets the human reviewer judge the work instead of excavating it.

That is the practical unlock. Not more babysitting. Not blind trust. A better shape for delegated work.

Give the agent a definition of done by task type. Require a trail. Keep production authority attached to named humans until the workflow proves it deserves more room.

The teams that win with coding agents will not be the ones that accept “done” fastest.

They will be the ones that make “done” mean something the business can actually live with.