AI-assisted delivery
AI-assisted delivery is useful when it produces evidence, not just output. These articles cover the receipts, provenance, and operational boundaries that make unattended coding work reviewable.
The path
32 min total
- Validation receipts: making “tests passed” tamper-evident
An agent that says the tests passed is making a claim about work you did not watch. A receipt binds that claim to a commit, a command set and the logs, so a wrong one is refused instead of believed.
13 min read - Provenance accounting: every commit on the branch explained exactly once
Receipts prove that a commit was validated; they do not say where it came from. Closing a pass of coding agents only when every commit on the branch has exactly one recorded origin is what keeps unexplained work out.
10 min read - What broke when I ran coding agents unattended for two months
The failures were rarely the model's code. They were the environment the agents inherited, the state they left behind, the boundaries I drew for them and the conventions they copied. Each one became a rule.
9 min read
Series in this path
- Parts 1 and 2 and 3 of Building a harness for coding agents