
The Three Types of Useful Agents (And Where They Win)
When people talk about AI agents, they often mean very different things. One way I find useful to separate them is by the job they do: coordinating work, making sense of information, or taking action. It’s an editorial lens, not a formal classification. Real systems mix the roles, and many jobs don’t need an agent at all.
1. Orchestration Agents
What They Do:
Orchestration agents help coordinate multi-step work: tracking what’s done, what’s next, and who’s waiting on whom.
Where They Can Help (illustrative examples):
- Project management: spotting a task that’s blocking several others and nudging its owner.
- Event planning: keeping track of dependencies between venue, catering and speakers, and flagging when one change affects another.
Where Ordinary Software Wins:
If the sequence is fixed, a checklist, a project template or a workflow tool with rules does the job more predictably. Coordination is also only as good as the information people put in; an agent doesn’t guarantee deadlines are met.
2. Insight Agents
What They Do:
Insight agents analyse information, look for patterns, and suggest what to look at or do next.
Where They Can Help (illustrative examples):
- Market analysis: reading a large pile of reports and pulling out themes worth checking.
- Decision support: assembling the relevant information ahead of a decision, with sources a person can check.
Where Ordinary Software Wins:
If the question is a well-defined number, a report or a database query gives the same answer every time. An insight agent’s output is a starting point for a person’s judgement, not a substitute for checking it.
3. Action Agents
What They Do:
Action agents carry out tasks, such as answering a question, updating a record or processing a request.
Where They Can Help (illustrative examples):
- Customer service: answering routine questions from approved information, and preparing straightforward requests for a person to approve.
- Routine operations: repetitive tasks with clear rules where mistakes are cheap and easy to reverse.
Where Ordinary Software Wins:
If the rules are clear, a form, a rules engine or a scheduled job will do the same thing every time, with no model involved.
Action is where limits matter most. An action agent should do only what it’s been explicitly authorised to do, with a named person reviewing its work and approving anything costly or hard to reverse, such as refunds, policy exceptions or commitments to customers.
A real case shows why. In Moffatt v. Air Canada, 2024 BCCRT 149, decided on 14 February 2024, British Columbia’s Civil Resolution Tribunal considered a customer who had relied on the airline’s website chatbot. The chatbot told them they could apply for a bereavement fare after travelling, which contradicted the airline’s own bereavement page. The tribunal found the airline liable for negligent misrepresentation. It rejected the airline’s argument that the chatbot was responsible for its own actions, finding that the chatbot was part of the airline’s website and the airline was responsible for the information on it (the tribunal’s decision, paragraphs 24 to 32). It’s a Canadian small-claims decision, not a general legal rule, and this isn’t legal advice. The practical point I take from it is narrower: from your customer’s point of view, what your agent tells them, you told them.
The roles overlap
Most useful systems mix these roles. A customer-service agent might pull together a customer’s history (insight), route the case to the right team (orchestration) and draft a reply (action). The lens is useful for asking which part of a job needs which kind of help, and how much authority each part needs.
It’s also worth asking whether you need an agent at all. Anthropic’s guide Building effective agents (December 2024) recommends starting with the simplest solution that works and adding complexity only when it’s needed, which can mean not building an agent. That’s vendor guidance. From the practitioner side, George Bezerra’s piece for CIO (August 2026) argues from his own experience for using deterministic rules wherever they’ll do, and making each step up in complexity earn its place.
When not to hand it over
Applied to the examples above:
- Orchestration: don’t let it reassign work or move deadlines on its own. Let it flag problems and propose changes for the owner to accept.
- Insight: don’t let its summary go into a decision unchecked. Someone checks the sources first.
- Action: don’t let a customer-service agent promise refunds, exceptions or policy outcomes unless those answers come from approved information, and have a person approve the costly ones. When it isn’t sure, it hands over to a person.
- Any role: don’t use an agent when a form, a rule or a report already does the job.
Before it acts
Pick the agent role you’re most tempted to build, and the one action it would take on its own. Before it takes that action, decide who granted the authority, what its limits are, who reviews its work, and what happens when it’s wrong. The autonomy test is a short way to write that down. For the broader question of what makes an agent worth building, see what makes an AI agent actually useful.
Updated:
Want help turning ideas like these into reality? Let's talk.