The intuitive design for an internal AI assistant is one agent that knows everything about the company. It is also the design that stops scaling at about the third use case.

One internal system we operate runs more than twenty agents, each responsible for a narrow slice, plus a handful of scheduled jobs that run without anyone asking. That split is not architectural decoration. Three forces push towards it.

One general agent Twenty narrow agents
Conflicting instructions guaranteed none
What prevents tool misuse the prompt the architecture
Blast radius of one fix the whole company one agent
Configuration to read when debugging one, very long one, very short

Instructions pull in opposite directions

Answering a leave-policy question well means quoting the rule precisely and refusing to improvise. Answering a customer question well means reading intent and filling gaps.

Put both in one instruction set and they dilute each other. The policy answers get vaguer; the customer answers get stiffer. Neither is wrong enough to notice in testing and both are wrong enough to matter in use.

Tool permissions are granted to the agent, not the platform

If a single agent must both read payroll and send external email, then that capability pair exists in your system and the only thing preventing misuse is the prompt.

Split it, and the guarantee becomes structural: the agent that talks to customers has no payroll tool, and the agent that reads payroll cannot send anything outward. No instruction has to hold the line, because the capability is not there.

This is why tool access should be configured per agent rather than per installation. A reader and a writer can be two different agents, and most of the time they should be.

Debugging stays local

Twenty narrow agents means a bad answer is traced to one configuration, and fixing it cannot regress the other nineteen.

With one general agent, every improvement is a change to the behaviour of the entire company's assistant. Teams learn this the hard way and then stop improving it, which is a worse outcome than the original bug.

The scheduled half

The part that changes how people feel about the system is usually not the asking. It is the jobs that run on their own.

Scheduled agents in that same system check, every day, what is due and unanswered, what does not reconcile, what has been sitting too long — and notify only when something is off. On a normal day they are silent, which is what keeps them from becoming noise.

An assistant you have to remember to consult gets forgotten in a busy week. A job that reports an exception does not need anyone to remember it.

Build order

Conversation first, schedules second.

A scheduled check needs a definition of "something is wrong", and that definition is usually not knowable up front. It emerges from a few months of people actually asking things — what they ask, where they get stuck, which anomalies turn out to matter. Specify the checks on day one and you will specify the imagined ones.