Agents should mind their own work
Giving one assistant access to every file, instruction, and connector may look like versatility, but it quickly becomes a liability. Reliability declines long before the failures become obvious.
The generalist assistant bottleneck
A single assistant can appear capable of handling everything. Give it enough instructions, files, and tools, and it can attempt almost any task. But every new responsibility also gives it more context to interpret, more rules to reconcile, and more ways to act.
The decline is gradual. The assistant may still produce plausible results while using the wrong source, applying instructions from another task, or changing something outside the intended scope. Nothing clearly breaks, but the work becomes harder to trust.
The problem is not that the assistant lacks intelligence. It is being asked to decide what matters across too many unrelated responsibilities.
Narrow the scope, improve reliability
Divide a larger objective into clear units of work. For each unit, give the assigned agent the instructions, files, tools, and permissions it needs—and no more than that.
A narrower scope makes the relevant context easier to identify and the result easier to evaluate. It also limits the consequences of a mistake: an agent working on one unit cannot accidentally interfere with unrelated work.
This does not mean agents should work in isolation. When one unit depends on another, agents can collaborate through explicit handoffs. Each agent remains focused on its own work while contributing to a larger outcome.