When to use AI agents (and when not to)
Agents are powerful and easy to over-apply. A practical framework for deciding when autonomy earns its keep.
Autonomous agents are the most exciting — and most over-applied — pattern in AI right now. Used well, they automate genuinely complex workflows. Used carelessly, they turn a reliable task into a flaky one.
Agent, or something simpler?
- The path is known and fixedA workflow, not an agent
- Steps depend on what earlier steps findAgent
- Actions are irreversible or high-valueAgent with a human gate
- One call answers itJust call the model
Agents shine when the path is uncertain
If a task has a fixed sequence of steps, you don't need an agent — you need a workflow. Agents earn their complexity when the path varies: when the system must decide which tools to use, in what order, based on what it finds.
Keep a human where the stakes are high
Full autonomy is rarely the goal. The systems we ship put humans in the loop at exactly the points that matter — approving an action, handling an edge case — while the agent handles the repetitive 80%.
Observability is non-negotiable
An agent you can't inspect is an agent you can't trust. Every action is logged, reversible and auditable. When something goes wrong — and it will — you need to see exactly what the agent decided and why.
Before an agent goes unattended
In place
- Every tool scoped and authorised independently
- Human gate on irreversible actions
- Full trace of every run
- A documented way to stop it
Not yet
- Permissions relying on the prompt
- No record of what it did last Tuesday
- Nobody owning it after launch
- Untested behaviour on hostile input