How I set up AI agents for complex, long-running work
I used to supervise my agents pretty closely, approving every change and reviewing in small chunks. These days I'll hand one a big feature and let it run for hours, and it mostly only stops when it hits a decision that actually needs me.
That isn't because I just started trusting it more. Three things changed: the models got a lot better at holding a long plan together, my tooling got better (AgentLink's Approve for Me mode has a Guardian reviewer that checks risky actions against what I actually asked for), and I put a lot more into the guidance up front.
The post covers how I set that up:
- Prep before the big prompt: research sessions, making sure the agent can run and verify what it's building, and skills for conventions it keeps getting wrong
- Writing a plan or RFC and getting it adversarially reviewed before any code
- Getting an early slice in front of me so I can steer early
- Breaking the output into reviewable PRs and shipping behind gates
The goal isn't a record-length run. It's longer stretches of useful work with fewer preventable wrong turns, while I still own the product decisions and what ships.
Keen to hear how others are running long sessions.






Comments
0 comments