What is a thin agent?

The analogy: a “thin client” for agents

If you have worked in technology, you know a thin client: a device that does very little on its own and leans on a server for the heavy lifting. A thin agent borrows the idea. It does one narrow thing — gather these facts, route this request, answer these three questions — and leans on a human for judgment. Small surface, predictable behavior, easy to trust.

Why thin is often the smart choice

• Smaller risk surface. Every capability you give an agent is another way it can go wrong. A thin agent has very few, so there is very little to go wrong.
• Predictable and testable. A fixed job with fixed inputs and outputs is easy to test, easy to explain, and easy to keep behaving the same way tomorrow.
• Easier to keep in-bounds. When an agent cannot advise, value, negotiate, or interpret rules, it simply cannot wander into the places that call for a human judgment.
• Faster and cheaper to ship. Narrow scope means less to build, less to manage, and a quicker path to something useful.

Where thin agents shine

Thin agents fit bounded, repeatable tasks where you want reliability and predictability more than cleverness:

• Intake and data collection — gather a fixed set of facts and hand a clean summary to a person.
• Routing and triage — sort an inbound request to the right human or queue.
• Status, scheduling, and qualification — small, repeatable jobs with clear right answers.

The design rules

• One job. Define the single task and refuse scope creep.
• Fixed inputs and outputs. A known list of questions in; a known summary out.
• No advice. The agent collects and routes — it does not advise, opine, value, or represent. That one rule does most of the work.
• Say what it is. Tell people clearly they are talking to an AI assistant.
• The human owns judgment. A person makes every call that matters, and handles the handoff.

A worked example

Imagine an intake assistant for a service business. It asks a short, fixed list of factual questions, records the answers, says a human will follow up, and does nothing else. It never tells the visitor what something is worth, never interprets a rule, never promises an outcome. The result is a tidy summary in a person’s inbox — useful, fast, and easy to keep honest. That is a thin agent doing exactly what it should: gathering the relevant facts and getting out of the way.

Thin vs. thick — choose scope on purpose

A thick agent — broad reasoning, many tools, real autonomy — can be powerful, but it demands far more: testing, logging, oversight, escalation paths, and careful governance. Neither is “better.” The discipline is to choose scope deliberately: start thin where you can, and only add capability when the job truly needs it — adding oversight to match.

Where to start

If you have a list of tasks you would like an agent to help with, thin agents are usually the smartest place to begin. Rank the candidates by how much they are worth, how often they repeat, and how ready they are to build — then start with the bounded, repeatable, low-risk ones. Automate the routine layer, keep a person accountable for anything that matters, and build the guardrails in from the start rather than bolting them on later.

Related: What is agentic AI? · The AI Compliant Twin · Thin agents & the Compliant Twin