A developer shares four workflow patterns for using AI agents effectively in software development.
Adapted from @milan_milanovic𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝘀𝗼𝗺𝗲 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗯𝗲𝘀𝘁 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀 Most developers today use AI agents for coding, yet we don't have any useful patterns for how to do it properly. What I see is that mostly everyone has their own style of working. I also improvised a lot, but after 3 years came to the point where I can see that some patterns emerge. Here they are: 𝟭. 𝗦𝗽𝗲𝗰𝘀 𝗳𝗶𝗿𝘀𝘁, 𝘁𝗵𝗲𝗻 𝗰𝗼𝗱𝗶𝗻𝗴 This is something I learned early, but also translated my pre-AI way of working. I first do specification or research and make plans with agents. Agents create SPECS.md or RESEARCH.md and PLAN.md docs. We do a few iterations here until I'm satisfied. Mostly I use AI agents plan mode for this. So the process is: Specify/Research -> Plan -> Tasks -> Implement. Then, I give the agent to implement, while I review results. 𝟮. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝘀 𝗶𝗻 𝘁𝗵𝗲 𝗿𝗲𝗽𝗼 During this process, I create docs (referenced from AGENTS.md and CLAUDE.md), so agents always know what we did and how. It also includes ARCHITECTURE.md where I note architecture decisions, and ADRs for architecture decision records. Then I set that agents should always update and review these docs during implementation, so they become a kind of living docs. 𝟯. 𝗧𝗲𝘀𝘁 𝗿𝗲𝘀𝘂𝗹𝘁 𝗼𝗳 𝗲𝘃𝗲𝗿𝘆 𝘁𝗮𝘀𝗸 So, when agents finish, I have a defined set of automatic tests, builds and screenshots (if UI is touched), where the agent checks its own work. Before that, I send it to another agent for a review. E.g. if I coded with Claude Code, I send it to GPT for a review. Different agents/models have different views on the code and can find potential issues. Also, I do this manually, check the code and UI. 𝟰. 𝗦𝗺𝗮𝗹𝗹 𝗯𝗮𝘁𝗰𝗵𝗲𝘀 𝗮𝗻𝗱 𝗳𝗿𝗲𝗾𝘂𝗲𝗻𝘁 𝗰𝗼𝗺𝗺𝗶𝘁𝘀 I instruct agents to work in small batches, because it's easier to do a rollback and I can also review it manually when needed. It also represents a history of changes that we can track later. All of these things, like specs, docs and small batches, are something we had before AI, but now it is more emphasized. The good engineering practices are here to stay, if you skip it you pay the price, maybe even bigger than before.
