Random list of tips (ymmv) from trying to get gpt 5.5 to build a big thing over a day or two:
- Use chunkier milestones with acceptance criteria rooted in end-to-end integration testing, not bullshit fake harnesses
- Have the root agent do the work with /goal and delegate to subagents to block on code review before every commit
- Make sure it's clear the plan is *immutable*, keep plan status updates in a separate file, don't let the agent cheat by updating the plan
- Let the agent add features and logging to the harness on demand to be able to test for acceptance, but ensure those also go through subagents for similar review to ensure they're necessary and not cheating
- Each milestone should have a dedicated folder with captured artifacts and distinct runner scripts so it can be fully audited for agent fuckery
- Part of acceptance criteria should be the end-to-end system is still possible to run interactively and be used by a user
- Have a cadence of "cleanup milestones" which force the agent to delete logging, debug code, kill dead code, remove all feature flags that are no longer needed, remove any 'fallback' or 'legacy' handlers, kill pointless error handling, and break up files and DRY up common code. imo don't do this at every milestone, do it every couple of milestones
- Make sure the plan includes detailed information up front of what the state shoudl be at the last milestone
- Include 'anti-patterns' in the plan - things the agent should never do, such as update the plan, or build a new harness with mock objects, or anything else that you discover in past runs as a loophole to not get shit done
- It's been helpful to have an interactive Opus session I can use as it runs to check the plan status and front run future milestones if things veer off track - instead of interrupting the primary agent, I have 4.7 tweak or insert milestones to course correct