1/ Some things I've learned recently running coding agents on large-scale projects. Most of this contradicts advice from 6 months ago!
2/ Think bigger. This is the most common mistake I see: tasks scoped too small. At this point you want to be aiming for work that would take a good engineer multiple weeks.
3/ Try using one long-running implementer session for the entire project. Mine routinely run for days or even weeks continuously, compacting many, many times. Compaction works now. Long sessions remember your conventions and patterns and you stop needing to re-explain things.
4/ Drive it with a persistent task list. Your job is to add vetted tasks faster than it can complete them. It's like shoveling coal into a steam engine. Each task should be: what to do, how to verify, proof notes once completed. The bar is: if satisfied as written, you'd trust
5/ Spend most of your time on plan docs, not watching the agent. I usually create plans in a short-lived planning session, append one or more tasks, then kill the session. Good plan docs are self-contained, specify interface-level details, and include an explicit end-to-end
6/ Adversarial review is what makes long, unattended runs possible. Before any task is checked off, a fresh read-only sub-agent reviews the diff against the todo + plan and return gaps. (This is often too powerful and you need to dial back from over-engineering).
8/ Set up long-lived sessions with roles: planner, implementer, adversarial reviewer, black-box tester, issue triager, deep code reviewer. Your job is to wire them together so the implementer is never idle, and monitor and review to make sure everything is working and catch
9/ Get yourself out of the loop. No manual PRs, no typing into the terminal, no checking CI. If you catch yourself testing for it, stop! The agent needs to prove the work is done; your job is to double check.
10/ Spend 20+% of your time on the meta. Make sure any mistake you notice is incorporated into future instructions so it doesn't happen again. Iterate on the process the agent is following. Improve the testing harness. However, be careful about over-engineering the workflow.