Originally published on LinkedIn.
I run FinMoment, an AI bookkeeping pipeline I built, with a team of 13 Claude Code agents. The biggest lesson from operating it: the model is not the part that needs the most engineering. The harness around it is.
By "harness" I mean everything that decides what an agent is allowed to do, how we know its work is finished, and where a human has to say yes. These are the rules I ended up with.
Rule 1: "Done" is a check, not a feeling
An agent will tell you a task is complete when it believes it is complete. That is not the same thing. Every task in my setup has a done-check: a concrete condition a machine can verify. The tests pass. The build succeeds. The change does what the task said it should.
If the check fails, the task is not done, no matter how confident the summary sounds. This one rule removes most of the "it said it fixed it, but it didn't" frustration people have with coding agents.
Rule 2: One job per agent, and a narrow set of tools
Thirteen agents sounds like a lot, but each one is small. A narrow role means a short instruction set, fewer tools, and a smaller blast radius when it gets something wrong.
It also makes failures easier to read. When something goes wrong, I know which agent produced it and which instructions to fix, instead of debugging one giant prompt that tries to do everything.
Rule 3: The pipeline is the boss, not me
I don't want to read every diff by hand. I want the pipeline to refuse bad work automatically. FinMoment runs on Next.js, Supabase and Vercel with GitHub Actions, and nothing ships unless the CI and deploy gates pass. Agents can propose changes all day; they cannot skip the gate.
This comes straight from my years building large systems (about 15, including fraud-decisioning platforms). Zero-downtime deployments only worked there because the pipeline, not someone's mood on a Friday, decided what shipped. Agents need the same discipline, arguably more.
Rule 4: Some decisions are human-only, by design
There is a short list of actions no agent can take on its own: anything that approves financial data, touches money, or changes what customers see in production. These are hard gates, not polite suggestions in a prompt. The agent prepares everything, then stops and asks.
This mirrors how FinMoment itself works. Uncertain fields go to a review queue, balances must reconcile, and a person approves before anything is exported. Nothing is guessed or auto-approved. The agents that build the product follow the same philosophy as the product.
Rule 5: Make the state visible
Agents lose context and sessions end. So the harness writes state where I can see it: what was done, what was checked, what failed, and what is waiting for a human. When I open my laptop I can see in a minute what happened and what needs me.
I use the same skeleton for smaller agents I run every day: one turns chat payment screenshots into a ledger and day book, one reviews daily farm pond logs and raises alerts, one hunts for market problems and keeps a tracker. Different jobs, same pattern: a narrow role, a done-check, a log, and a human gate for anything that matters.
Rule 6: Treat agent access to your tools like an API key
FinMoment exposes an MCP endpoint so Claude can drive it. That is powerful, and it deserves the same care as any credential: scoped access, row-level security in Postgres, and human approval before anything important is committed. An agent connected to your tools is a new user of your system. Design for it like one.
What this means if you're adding agents to your business
You don't need 13 agents. You probably need one or two, with clear done-checks, a pipeline that blocks bad work, and a short list of decisions only a person can make. That is where most of the reliability comes from, far more than from picking the "best" model.
That's the work I do: AI automation for finance and document workflows, Claude agents and agent harnesses with approval gates, and AWS/platform architecture. If you'd like to see a harness like this running on a real product, I'm happy to walk you through FinMoment on a call.
You can find me on Upwork: https://www.upwork.com/freelancers/~0146d40003db2cffd3