I gave three coding agents less capability on purpose

Claude Code is built around a general-purpose agent that can do almost anything in a codebase to help ship faster. I customized 3 specialists to run beside it, each deliberately able to do less.

Why it matters

Hire a generalist and they can do anything. That is the value and the risk: nothing tells them which calls to escalate or when to stop.

An agentic coding tool is the same. It reads your codebase, edits files, runs commands, manages version control. For open-ended work, that reach is the point.

When the work touches systems where mistakes cost money, capability stops being the constraint. Judgment does. The agent has to know what to refuse and when to stop for a person.

That judgment is the layer I add on top of a tool that already works.

Much of the engineering work in agentic AI lives in the boundaries.

How it works

A Principal Investigator, a Lead Engineer, a Staff Reviewer. 46 explicit prohibitions between them.

Some boundaries are structural. The Engineer’s tool list holds no web and no research tools, so it cannot quietly guess an API. It consults the Investigator by name, with the exact library and file line, which turns a vague hand-off into a request someone can answer.

Others are prompt rules, and I keep that difference honest. Anthropic documents that the agent-spawn allowlist binds only the main session agent, so I do not call the pipeline sealed.

The Reviewer re-runs all 8 quality gates itself instead of reading the Engineer’s report. A claim is not evidence.

The fair objection is that gates are friction, and every one of them is maintenance debt. Mine sit only at the steps that are expensive to undo, which is the test I apply before adding another.

Every boundary above cost something to add. The useful question to put to anyone running a setup like this is not which boundary they added first. It is which one they would remove first.