Nothing Left to Audit

A paper called The End of Software Engineering went up on arXiv on 4 June 2026, and one person signed it. Zhenfeng Cao, of Lingxi Intelligent Investment in Shenzhen, an investment company rather than a lab. On X it became "Chinese researchers published a paper", and the plural carries an authority the affiliation does not, which is a question about who we are actually listening to. The paper is worth reading anyway, and not for its title.

The argument is clean and built the way an engineer builds one. Three generations of software delivery: installed locally, hosted in the cloud, and now the agent. Each shift moves complexity one step along, from the user to the vendor and from the vendor to the model. In the third, code stops being the work and becomes equipment, generated inside a reasoning loop, executed, thrown away. Cao states it plainly: "What persists is the agent's capability, not its intermediate artifacts."

That sentence also names what stops persisting. Permanent code was the record of a decision: someone reads it, diffs it against last week's version, traces who wrote it, and argues with the choice line by line. A loop that writes and discards keeps the decision and drops the record.

The same paper promotes the human and demotes the evidence in a single move. The developer becomes an intent architect, an agent coordinator, an outcome auditor. An audit needs an object that outlives the run. The document handing you the title is the one designing the disappearance of the thing you would audit. Nobody wrote that contradiction on purpose. It falls out of the incentive: the artifact is expensive to keep and the capability is what gets sold.

Cao's own numbers show where that record would matter most. The EvoClaw benchmark he cites measures agents in two regimes: above 80% on isolated tasks, at most 38% under continuous operation, a 54% drop. Chained actions fail more often than they succeed, and that is exactly the condition in which you want to reconstruct the step where the chain bent. It is also the condition in which the paradigm leaves nothing to read.

A second figure in the paper points the same way. His showcase result is a coordinated agent system that cut root-cause identification time by 93% and saved more than 200 engineering hours in a month. The strongest win of the new paradigm is debugging, which means working out why something went wrong. That trade runs on traces, and it is being celebrated inside an architecture that makes traces ephemeral.

Of course the logs exist. A serious agentic system records prompts, calls, tools invoked, outputs. A log answers what was done, while source code answered how the choice was organised: the structure, the cases covered, the cases ignored, the point where someone decided two conditions were enough. That form is something another person can contest.

Custody decides what an investigation actually meets. Ask a vendor to reconstruct a decision from eleven months ago and the answer arrives shaped by a retention window you did not set, a schema you did not design, and a redaction policy written for the vendor's exposure rather than yours. The failure mode is familiar to anyone who has run a postmortem against a third-party service: the data exists in principle, arrives in a form that cannot be diffed, and lands well after the moment when it would have changed anything. Source code sitting in a repository you control has none of those properties, and that is the property being traded away.

This is where the word auditor changes weight. Signing off on an outcome nobody can reconstruct produces a receipt, and a receipt shows that something happened, never who answers for it.

The job Cao describes is arriving, and the part worth negotiating is not in the title. If your work becomes specifying intent and validating outcomes, your protection is whatever survives the loop: which artifacts persist, how long they are kept, and who holds them. Those are three questions for a contract, not for a thread.

Cao is right about what persists. The agent's capability stays, and it stays with the vendor. The record, the one you need, stays with no one.