01Start HereThe map + Core 10 02What AI Still Gets Wrong11 failure patterns 03All 30 TermsSearch the guide 04What Is an Agent?Agency, tools & autonomy

What Is an Agent?

Agency, tools & autonomy.

An agent is more than a model thinking through an answer. Agency begins the moment reasoning starts controlling actions — reading the results, and deciding what to do next in pursuit of a goal.

Agency is a spectrum, not a switch.

It can start with a single, bounded tool call and grow into hours of independent work across real systems and real data — with permissions and checkpoints shaping every step along the way.

More human-directedMore model-directed
01 Seconds

Reasoning

The model works through alternatives and constraints before answering.

Still just an answer.
02 Seconds → minutes

Tool Use

The model searches, calculates, or takes one bounded action.

Capability reaches past the answer.
03 Minutes

Workflow

A defined sequence coordinates AI, rules, checks, and handoffs.

The path is mostly fixed in advance.
04 Minutes → hours

Agent

The model chooses its own actions, watches the results, and adjusts toward a goal.

It helps decide what happens next.
05 Potentially hours

Multi-Agent System

Several agents take distinct roles and coordinate or check each other's work.

Coordination adds capability — and complexity.

The key distinction

An agent plans, acts, observes, and adjusts.

Reasoning by itself stays inside the answer. An agent uses that same reasoning to decide what to do next in the world — inside whatever tools, permissions, checkpoints, and stopping rules it's been given.

A fixed workflow might always run the same three validation checks, in the same order, on a data extract. An agent asked to reconcile that same extract against a source system can decide for itself which records to re-check and when the discrepancies are worth flagging — a meaningfully different risk profile, and a different governance conversation.

01Plan
02Act
03Observe
04Adjust

Anatomy of an agent

Five questions worth answering before anything gets deployed.

01

Goal

What outcome is it actually pursuing?

A vague goal produces an agent that solves the wrong problem efficiently. The clearer the definition of success, the easier everything below is to govern.

02

Model

What's doing the reasoning and choosing?

The model's own capabilities and blind spots set the ceiling for everything built around it — no amount of tooling fully compensates for a model that reasons poorly on your specific domain.

03

Tools

What can it read, calculate, or change?

This is where the real risk profile gets decided. Read-only access to a report is a very different conversation than the ability to update a live record.

04

Guardrails

What's explicitly off-limits?

Permissions, approval gates, and rate limits that hold even when the agent's own judgment points somewhere else.

05

Human Judgment

Where does a person have to check in?

The checkpoints that catch a mistake before it reaches a patient record, a public report, or a signed contract — placed where errors are expensive or hard to reverse.

Evaluating an agentic AI proposal?

We help health authorities scope agent permissions and guardrails before a pilot goes live — not after.

IMC holds no allegiance to any AI platform or vendor, so the read you get on how much autonomy a proposal is actually asking for is independent, and grounded in what your organization can safely govern.

Start a conversation