Scoping AI for teams
Everyone is building
one AI that
does everything.
The teams pulling ahead
run smaller ones.
Most teams answer AI mediocrity with a bigger assistant, more tools, more context. That is the wrong direction. Anthropic's own guidance opens with a warning: start with a single narrow agent, not a stack of them. Here is the pattern, plus how one client ran a full audit through five of them and knew exactly which one broke.
The argument
An agent given less to do does better work.
The dream is one assistant that writes, analyzes, checks the brand, books the meeting, and remembers it all. In practice it is a generalist holding too many jobs, giving none of them full attention. Context bleeds between tasks. Tools get grabbed wrong. What looked sharp on one job in the demo goes mushy the moment it has to hold five.
The move is narrower agents. More of them is beside the point.
The pattern
One job per agent. A coordinator holds the map.
A human owns the checkpoints.
Give each piece of work its own agent with one job, one clean set of tools, one clear boundary. The scoping is the skill. The machine will happily take on everything and get worse without telling you.
One job. One toolset. One boundary.
Not a generalist with limits. A specialist scoped to a single task, with only the tools that job needs. Take capability away and output gets better, because the agent stops deciding what to work on and does the one thing.
Holds the whole picture.
One agent that maps the work, routes to specialists, and knows what came back from each. The generalist role belongs here, at the traffic level, not inside the workers.
Owns the exceptions.
Narrow agents relocate the person. Instead of babysitting one overloaded assistant, a human sits at the checkpoints where judgment is worth the most. McKinsey puts that judgment at 70 to 75 percent of the variable cost of an agentic workflow. Design for it.
The proof
A podcast audit, run through five small agents.
One client, one workflow, five narrow agents plus a human at the end. When something broke (and something always broke) I knew exactly which agent missed and fixed that one piece. Try that with one giant do-everything prompt.
A giant do-everything prompt fails like a black box. A line of narrow agents fails at an address you can find.
Built with you
The Written Brain
Setup Sprint
splits one agent, too.
Five days,
$4,500,
yours forever.
The framework on this page is free. Take it. If your AI gives impressive demos and mediocre daily work, though, the setup is a judgment call and a design call. That is the Sprint.
What ships in five days.
- One over-scoped agent on your stack, taken apart and re-scoped as a narrow-agent line.
- The coordinator, wired to route between the workers.
- The Written Brain files (state.md, decisions.md, knowledge.md) so the split holds across sessions.
- Human-gate checkpoints at the places judgment actually lives.
- A two-hour handoff so your team owns the setup, not me.
Send this to someone
If someone you know keeps loading up one giant assistant and wondering why it slips.
Copy the link. Send it to them. The framework is on the page, no download, no gate. Also lives at jeffschenck.com/narrow-agent.
Start
If your AI gives
impressive demos and
mediocre daily work,
look at the scope before you
blame the model.
Bring the workflow. I will tell you where I would split it.