The first week

The quickstart is twenty minutes of setup. This is the week that follows it, which is the part that decides whether the thing gets used. It is written from what the product does rather than from a rollout methodology, and the only instruction that really matters is the one on day one: pick a job small enough to be boring.

Day one — one connection, one agent, no ceiling

Connect the system where your orders arrive. Brief one agent on one repeated job. Leave its spending ceiling at nothing, which is where a new agent starts, and route everything you can to the approval queue. The agent will read, classify, draft and hand off; it will not spend. That is the correct first day.

Days two and three — read the traces, not the board

The board tells you what happened. The trace tells you why, and in the first week why is the only thing worth your attention. Open three or four runs in full and read what the agent looked at before it decided. Two things usually show up: a brief that is more ambiguous than it looked, and a source the agent keeps consulting that turns out not to be authoritative.

Day four — rewrite the brief once

By now you will have seen the agent do something reasonable that you did not want. Almost always the fix is one of the two sentences most briefs are missing: what to do when unsure, and what never to do. Add them, and leave everything else alone so you can tell what the change did.

Day five — raise the ceiling by a small amount

Set it to a figure you would shrug at losing over a week. Not a figure that covers your busiest day. The point of the first raise is to see what the agent does when spending is possible at all, and a small number tells you that just as well as a large one.

Day six and seven — the first brief that matters

After a full week the daily brief has something to compare against. What settled, what stalled, what waited on you and for how long. The number to look at is the oldest waiting item: if something sat in the approval queue for fourteen hours, the queue is not being read, and no amount of agent capability fixes that.