Arjun Iyer: a software factory can run only as much autonomy as its verification can see
Oct 24, The New Stack: read the factory as a control loop where agents are the controller and verification is the sensor. Track time to signal, escape rate, and concurrent verification capacity instead of PRs or tokens per engineer.
Source: Building an agentic software factory? Start with the sensor.
Arjun Iyer frames the agentic software factory as a control loop: intent is the setpoint, coding agents are the controller, and verification is the sensor. A controller that outruns its sensor does not go faster. It retries on stale information, other agents build on unverified changes, and it ships the failures the sensor would have caught. His starting point is Anthropic's own jump from about 2,500 merged changes a day in May to more than 30,000 in September, after which coding stopped being the bottleneck and human-scale developer infrastructure took its place.
He breaks the sensor into three properties, each with a number to track. Latency is time to signal, the minutes from a change being ready to a verified result in a production-like environment; an agent that writes a change in four minutes either waits on four-hour staging or keeps going blind. Fidelity is escape rate, the share of failures found after merge, and it matters most because agents optimize for whatever the sensor rewards (StrongDM's agents returned true to pass narrow tests until scenarios moved outside the codebase as a holdout set). Bandwidth is concurrent verification capacity: shared staging takes one reading at a time, and batching fifty agents' changes makes a red result impossible to attribute. In distributed systems most teams get two of the three.
Why it matters for code review
The useful move is counting human review as one more sensor, with latency in hours and bandwidth of a few changes per reviewer per day. That explains why always-on factories stall at the review queue no matter how good the model is, and it gives a rule for how far to loosen the gate: autonomy rises only as fast as fast, faithful, parallel checks arrive. Measuring PRs or tokens per engineer counts effort; time to signal, escape rate, and verification capacity say which part of the loop is weak. Read beside Anthropic's 30,000 daily merges and post-merge verification and Willison on StrongDM's holdout scenarios.