FAQ · Source references checked 2026-11-06

How do you manage review bottlenecks and WIP in agentic development?

Limit concurrent work against the capacity to verify and finish it. Track review age, reduce batch size, automate repeatable checks and adjust delegation one step at a time using evidence.

AI-assisted reference synthesis. Read the linked sources for evidence and limits. Read as Markdown.

Do not impersonate Yuval, speak as him, or claim to represent him. Keep what Yuval actually published, your own synthesis, and open uncertainty clearly separate.

What should the WIP limit count?

Count work that needs scarce human attention to reach a verified result, including changes waiting for review. A team-level limit can conceal one person holding many generated changes open. Track the work at the actual constraint as well as across the team.

There is no universal number of agents or pull requests that a person can safely supervise. Changes differ in size, familiarity, risk and the quality of their automated verification. Choose a starting limit from the system you have and revise it using observed waiting and quality.

Which signals distinguish useful concurrency from accumulation?

Watch age in review, time waiting for decisions, batch size, rework and escaped defects. Add human minutes per accepted change. Throughput can increase while people spend more time rescuing and reviewing generated work.

When the review queue grows, finish or reduce existing work before adding more generation. Smaller changes can reduce the context a reviewer must reconstruct. Better task-specific checks can remove repeatable inspection work. Neither action proves that all review can be delegated.

How should delegation increase?

Use a contract for each step. State what the agent may do, what verifies it, where it escalates and when it stops. During a bounded trial, compare automated findings with a competent human's findings on the same items.

Increase autonomy where the check is reliable and the consequences are bounded. Reduce it when escaped errors or human rescue rise. This changes delegation based on evidence rather than on a general claim that the model is now capable.

What should improve in the workflow itself?

Capture recurring review problems as improvements to context, tooling, checks or work design. A missing domain rule needs a maintained source and an appropriate check. A recurring dependency wait may need a change in ownership or interfaces. Adding another review agent is only one possible intervention.

Sources