The Audit Trail Vanishes When AI Ships Around the Clock
Agentic coding collapses the delays that gave engineering oversight its foothold, and the test reports that replace them come from the system being audited.
Agentic coding collapses the delays that gave engineering oversight its foothold, and the test reports that replace them come from the system being audited.
When AI coding agents plan, write, test, and deploy in parallel, the thing that disappears first is not the engineering team. It is the time the team used to have. The delays between commits, between code review and merge, between a release and the next morning's incident report, are not wasted hours. They are the medium of accountability, and the agentic production loop destroys them.
That is the warning running through a recent essay by practitioner Olivier "owulveryck", the second in a series on what he calls the agentic platform: the systemic context, guardrails, and tooling that let AI agents ship code on their own. Where part 1 defined the what, this post tackles the who by mapping the agents onto Matthew Skelton and Manuel Pais's Team Topologies framework. The argument is structural, not just operational. Traditional software production is time-distributed for a reason. It gives humans room to question, challenge, and gate the work. Agentic production compresses that work into a continuous, parallel loop, and the gatekeepers have to be re-placed somewhere else.
The cognitive-load reshuffle. Skelton and Pais built Team Topologies around cognitive load: the idea that team boundaries exist to keep any one group's working memory from overflowing. Four team types, stream-aligned (product delivery), platform (internal services), enabling (coaching), and complicated-subsystem (specialist expertise), are matched to three interaction modes (collaboration, X-as-a-Service, and facilitating) so the seams between teams do not become seams in the system.
Owulveryck's move is to point out that an "agentic factory," a production loop in which agents plan, code, test, and deploy with minimal human in the loop, does not reduce cognitive load. It moves it. All the judgment that used to sit in design reviews, code review, QA, change advisory boards, and compliance checks now sits in one prompt-sized window on one human's screen. That is the wrong shape for that load.
In a discussion on Hacker News, the post has drawn only light engagement so far (five points within hours of posting), which suggests this is still a niche conversation. The audience for it, though, is not niche at all. Any organization that needs to remain governable while shipping software faster than a human can read it is in scope: banks, insurers, regulators, public-sector IT.
Why the existing audit trail stops working. The deeper problem, and the one the post hints at more than it spells out, is what disappears when production speed outruns review speed. The audit trail in a traditional software shop is built from artifacts that arrive in time. A pull request sits for a day. A release candidate waits for sign-off. A deployment window has a cutoff. A post-incident review has a date. The "trail" is not the log file. The trail is the spacing.
When agents ship continuously, the spacing collapses. Commits land in seconds. Tests run as the code is written. Deployments happen on green. The only artifacts left to inspect are produced by the same agent, or the same fleet of agents, that wrote the code. The "evidence" of a clean ship is a report from the shipper.
This is not a tooling problem. It is not solved by adding a better dashboard, a stricter linter, or a more detailed deployment log. Those are themselves outputs of the same system being audited. The institution designed accountability into time-delayed gates, and the agentic factory removes the gates and the medium with them.
What the new gate looks like. Owulveryck's answer is that the platform team becomes the gate. The platform that supplies the agents with context, guardrails, and tools takes on the anticipation work that the human gates used to do: policy-as-code, security review baked into the tool schema, test coverage enforced before an agent is allowed to write. The "who does what" of the Team Topologies mapping is that the platform team is no longer a service provider to the agentic factory. It is the institutional memory the factory cannot have.
Three things follow from that, and they are worth watching in any organization that adopts this pattern seriously.
First, the security boundary is now the tool schema. An agent that cannot call a destructive function cannot be tricked into calling one. The prompt-injection and tool-misuse risk that has been a recurring concern across the AI infra stack does not get solved in the prompt. It gets solved in the platform's interface design.
Second, the regulator's legible artifact changes. Compliance frameworks that assume a pull request, a reviewer, and a release date will not, on their own, disappear. They will be replaced, slowly and awkwardly and usually after incidents, by artifacts the platform can produce: signed policy bindings, capability grants, full action graphs.
Third, the team topology has to be designed, not inherited. A "swap the tool, keep the org" approach, which is the default for most enterprises adopting AI coding, will hit a throughput ceiling that looks like a model problem and is actually an organizational one. The ceiling shows up as agents waiting for human approval, or humans approving things they did not read, or audit findings after the fact.
The part to watch. The piece that owulveryck's series is building toward is whether the platform team can actually absorb the anticipation burden. The Team Topologies framework assumes platform teams are a small, senior slice of the org. In an agentic factory, they become the institution's primary risk surface. That is a very different kind of team to staff, govern, and keep from becoming the next single point of failure.
The load-bearing question for any executive weighing this transition is not "how fast can our agents ship." It is "what in our current oversight model is actually load-bearing, and what is just the texture of the time we used to have." Once that distinction is made, the platform team has a real job description. Until then, the audit trail is the schedule, and the schedule is gone.