Thin agents and the Compliant Twin: narrow scope as a guardrail

The risk-surface idea

Think of everything an agent can do as its surface area. A broad assistant that can reason about anything, pull many tools, and speak with authority has a large surface — and a large surface is hard to keep predictable. A thin agent, by design, can do almost nothing except the one job you gave it. There is simply far less that can go wrong, so there is far less to manage.

The one rule that does the heavy lifting

For most narrow use cases, a single design rule carries the load: the agent collects or routes — it does not advise. Paired with a clear disclosure that it is an AI assistant, and a human who makes the judgment calls, that one rule keeps the agent doing only what it was built to do. It does not opine, does not value or promise, and does not stand in for a person — because it was never built to.

The Compliant Twin connection

An AI Compliant Twin is the idea that your public-facing AI should face the world inside clear boundaries, with a keeper in the loop — a person accountable for what it does. A thin agent is that idea at its smallest and most reliable scale: the thin agent is the friendly front door that gathers and routes; the person behind it is where judgment and decisions actually live. Thin in front, human behind.

Designing a thin agent

• Define the single job — one task, written down, with scope creep explicitly off the table.
• Fix the inputs and outputs — a known list of questions in, a known summary out.
• Make “no advice” a built-in rule — the agent gathers and routes; it does not opine, value, or stand in for a person.
• Say what it is — tell people plainly they are talking to an AI assistant.
• Define the handoff — how and when a human takes over.
• Keep data minimal — collect only what the job needs.

When you outgrow thin

Sometimes you genuinely need broader reasoning. The rule of thumb is to add capability deliberately and add oversight to match — testing, logging, and human review grow alongside the agent’s reach. Capability without matching oversight is where trouble starts; thin agents avoid that by simply not reaching.

Keeping it simple

The habit that makes thin agents work is restraint. Give the agent the smallest job that is genuinely useful, keep a person responsible for anything that calls for judgment, and only widen its scope when a real need justifies it. Thin where you can — and let people, not the agent, own the decisions that matter.

Related: What is a thin agent? · The AI Compliant Twin · What is agentic AI?