The interesting part is what stood between agents and main
Cockroach Labs reports that its MOLT Sinai workflow processed more than one million lines of code in just under five months. The company says the system produced 1,238 pull requests and seven reverts. Its agents used separate roles for workup, plan review, implementation, code review, and the final merge audit.[1]
Those numbers are company-reported results, not an independent evaluation. The authors also say the first IBM Db2 experiment ran before human approval mode and that they were still confirming whether that support was correct. Later customer-bound work required human review before merge.[1]
The gate matters more than the costume.
Plan review moves failure to the cheap side of the diff
MOLT Sinai requires a reproduction, diagnosis, change plan, test plan, risks, and open questions before implementation. A separate agent reviews that plan. If scope changes during the edit, the agent stops and returns to plan review instead of improvising.[1]
This is the reusable part. A role name can help a model stay consistent, but the control comes from an inspectable transition. The plan has named fields. Another worker can reject it. The implementation cannot quietly widen its own scope.
A handoff needs read-back
The workflow borrows I-PASS for stuck work. The outgoing worker records severity, summary, actions, expected risks, and a receiver synthesis. The receiving worker writes back what it understood before continuing.[1]
BMJ Quality & Safety describes that synthesis step as closed-loop communication. It gives the receiver a chance to expose a mistaken understanding before acting. The evidence behind I-PASS is broader and more mixed than one headline number. A large multicenter study reported lower error and preventable-adverse-event rates after implementation, while a later cluster trial improved handover compliance without changing its primary adverse-event outcome.[2]
Only one role should change the system
Google's incident-management guidance separates command, operations, communication, and planning. It says the operations team should be the only group modifying the system during an incident. It also requires a live state document and an explicit handoff that the receiver acknowledges.[3]
Agent workflows need the same ownership. A reviewer can inspect. A planner can propose. One implementation role owns the edit. The merge role checks that review and test evidence exist. If every role can patch the branch, the chart no longer explains who changed what.
Review loops need a brake
Cockroach Labs reports one small fix taking eleven rework rounds over two days. Its later rule sends work to an attending after four consecutive rounds that each find a new blocker, or six substantive rounds of any kind. The authors say the rule was recent and did not yet have enough data to show its effect.[1]
A count is not a quality score. It is a stop trigger. Once the trigger fires, a new reviewer should inspect the goal, plan, evidence, and unresolved disagreements. Spending more tokens inside the same loop is not escalation.