18 MAR 2026
David's approach: Iterative multi-agent plan review before execution
David's standard pattern when building a complex plan is to launch multiple independent sub-agents to audit it from different angles, then update the plan, then audit again — repeating until confident — before allowing any code to be written.
When finalising Kerra's Canvas file metadata heal and external source recovery plan, David didn't proceed to implementation after a single review pass. Instead, he ran iterative parallel auditing loops:
okay, review the plan, go and launch 6 gpt 5.4 xhigh sub agents to review it as well, and tell them to then go and audit the codebase end to end to understand how it works, identify any issues with the plan, any lingering failure modes the system still has even after the plan is executed, any race conditions or failure modes or edge cases the plan will introduce
After the sub-agents returned findings, David directed the primary agent to:
dive back into the codebase and the sub agents outputs and understand every issue they raised, then dive back into the codebase to understand them more and what the solution needs to be for them, then you need to update the plan based on each of the sub-agents finding, solving for each issue that arose. then, once the plan is updated, then you should review the new updated plan, and then once again launch another set of 6 gpt 5.4 xhigh sub agents to review it as well
For the execution phase, David applied the same discipline: orthogonal, independent work can be partitioned to parallel sub-agents, but any work requiring cognitive coherence must be done by a single agent to avoid contradictions.