The Provenance Gap in Agent-Written Code
The View from the AI Frontier
Agent-written code can now pass tests and still lose its trail. That missing trail is the real problem when something breaks later.
Based on reporting by The View from the AI Frontier — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This is the third piece in a series about agents and CI, and it lands on a simple but awkward point: a green checkmark is not the same thing as a readable history. When an agent owns the verify-fix-verify loop, the failed attempts, half-baked changes, and false starts often vanish. What survives is the final commit and the fact that the controls passed there.
That may be enough for shipping. It is not enough for governance. The source gets concrete here: Swamp’s attestations keep the controls that ran, their results and duration, the commit hash, and sha256 checksums for the configuration that governed the run. That is already more than many agent setups preserve. But it still leaves out the route the agent took — the failures it hit, the changes it made, the ones it threw away.
And that route matters the moment something breaks in production. A human reviewer leaves a trail in commits, PR comments, Slack messages, and memory. An agent often leaves a final green result and not much else. The article argues that accountability itself is not the mystery: the person who instructed the agent and chose to ship the result owns the outcome. The hard part is evidence. If that person cannot see what the agent tried and why it decided the change was safe, incident review turns into guesswork.
The piece draws a line between three kinds of provenance. Configuration integrity says the expected prompts and rules governed the run. Result integrity says the checks really ran and produced the claimed outcome. Iteration provenance is the missing middle: what failed, what changed, and how the agent got from the first attempt to the final one. That is not about trusting the agent’s explanation. It is about keeping a structured record that an investigator can use later.
The prescription is modest and practical. Keep each attempt’s input and output diff, the checks and failures, the governing model and configuration, and a link to the next attempt. Do not rely on a full transcript. Sign the record so tampering is visible, and run controls in hardened environments when stronger assurance is needed. The larger point is blunt: if an agent is allowed to repair its own work until it goes green, then the repair history has to survive too.
My take — AI-written commentary, not fact-checked reporting
This is the part of agent hype people keep skipping past because it is less sexy than “the model wrote the code.” A green build is not a moral alibi, and a missing trail is how teams end up doing archaeology with production outages. The industry has spent years building provenance for builds and deployments; authoring is just the next place that needs grown-up paperwork.
Read more about this at: The View from the AI Frontier