Play 4: Put Agents on Narrow Jobs
Broad context, narrow skills, people on the decision
Play 4: Put Agents on Narrow Jobs
Goal: agents take on a handful of repeatable jobs end to end, and every decision they prepare goes to the person who owns it.
Pick the first jobs
Choose three to five jobs that are repeated often, painful, well understood and low-risk to get slightly wrong. Good candidates sit in the operating rhythm of the business:
- Collecting status. Ask owners for updates, chase what's missing, write the answers back to the model.
- Explaining variance. Trace the gap between plan and actual to the initiatives and decisions behind it.
- Preparing reviews. Draft the readout for a steering meeting, a portfolio review or the board from the model.
- Checking dependencies. Find the work at risk because something it depends on slipped.
- Gathering control evidence. Find which controls lack evidence and ask their owners.
Write the job down as a skill
A skill is a written-down method: the question it answers, the steps it follows, the records it reads, what it produces and who approves what it proposes. Writing it down does three things: it makes the job repeatable by anyone allowed to run it, it makes the agent's behavior reviewable, and it makes improving the job a matter of editing a document rather than retraining a person.
Most teams can now draft a skill by describing the job to an AI and refining what it proposes.
Keep people on the decision
- Draft, then confirm. Agents propose changes; people apply them.
- Route approvals to the owner. A proposed budget move goes to the budget owner, not to whoever ran the agent.
- Run with the approver's authority. When an approved change is applied, it should run within the permissions of the person who approved it.
Outputs
- Three to five skills in regular use, each with an owner.
- A list of the decisions each skill prepares, and who owns them.
Signs it's working
- Reviews start with the facts already assembled.
- The people who used to do the assembly are doing the analysis instead.
Pitfalls
- Starting horizontal. "An agent for everything" is how pilots die. Start narrow.
- Hiding the method. If nobody can read what the agent does, nobody can trust it or improve it.
- Automating the decision. The value is in preparing decisions, not taking them away from accountable people.