Day 137: One Agent Per Job
If I picture a normal ten- or fifty-person company using Pinchy well, the picture is never one genius assistant that does everything. It’s a small staff of narrow agents, one per job: a bookkeeper that lives in the accounting system, a mailbox agent that drafts the replies nobody wants to write, a support agent that actually knows the docs, an onboarding guide for new hires. Each one is dumb outside its lane and sharp inside it. That framing, one agent per team function, is the version of “AI at work” I think actually survives contact with a real org, and it’s a deliberate positioning choice, not a UI accident.
Positioning with an engineering bill
Choosing “many narrow agents” over “one broad assistant” sounds like marketing until you notice it sends you the bill in code. The framing only works if two things are true, and both are load-bearing.
First, adding a job has to be cheap, or you’ll never have more than two agents and the whole picture collapses back into one overloaded assistant. That’s why integration plumbing has been the theme lately: every new job an agent can do is usually a new system it needs to reach, and if reaching each system is a bespoke hack, the roster stays tiny.
Second, each agent’s edges have to be tight and legible, or a roster of narrow agents is just a wider attack surface wearing uniforms. A bookkeeper that can also read the whole team’s mailbox isn’t a bookkeeper, it’s a liability. This is the fail-closed permissions argument again, from the org-chart angle: narrow only means something if the narrowness is enforced, not just intended.
Why the plumbing weeks matter
Both of those showed up as this week’s supposedly unglamorous work. The reason Penny needed her lookups scoped to a single company is the same reason a mailbox agent has to sit behind one provider-agnostic contract instead of a Gmail-shaped assumption baked into everything: a job-shaped agent needs its data access to be exact, and exactness is a property of good boundaries, not of a clever prompt. The next few days are more of the same groundwork, widening which mailboxes a mail agent can speak to and cleaning up what happens to an agent when its connection is deleted and re-added, so that staffing a function with an agent is a boring, safe, repeatable act rather than a one-off integration adventure.
Day 137
The dream everyone sells is one AI that does it all. The thing I’d actually want running my own company is a handful of agents that each do one thing and can’t do the others, on infrastructure I run myself, where I can see exactly where each one’s edges are. That’s less impressive in a demo and far more trustworthy on a Tuesday. Most of the recent work looks like plumbing because it is plumbing, and the plumbing is what decides whether “one agent per job” is a slogan or a thing you can actually staff.