Learn › Agent Harness, the big picture › Stop 2
Stop 2 of 9 · Page
Autonomous agents
What autonomous agents are, the five levels of autonomy, why businesses want them, and the catch.
What it is
An autonomous agent is an agent that nobody is watching step by step. It starts on an event or a schedule, works for minutes or hours, and acts by itself inside limits that someone set in advance.
Autonomy is not a fourth kind of application. It is a second dial. The kind (stop 1) says who decides the next step. Autonomy says how much a person checks.
| Level | What the AI does | What the person does | Example |
|---|---|---|---|
| 0 · Answer | Answers one question | Reads it and decides | Summarise this email |
| 1 · Suggest | Proposes an action | Takes the action | Drafts a reply for a claims handler to send |
| 2 · Act with approval | Acts, but asks before anything important | Approves or rejects | Asks before making a refund |
| 3 · Act within limits | Acts by itself inside a budget and a set of rules | Reviews afterwards | Settles small claims and logs every step |
| 4 · Run unattended | Starts itself, runs for hours, recovers from problems | Sets the goals and limits, reads the report | Handles incidents overnight |
Why businesses want them
Each benefit below comes with the number you would measure to prove it. A benefit you cannot measure is a claim, not a benefit.
| Benefit | In an insurance claims team | How you would prove it |
|---|---|---|
| Work does not wait for office hours | A claim reported at 11 pm is checked and acknowledged at 11 pm | Time to first response |
| Faster from start to finish | Small, simple claims settled in minutes, not days | Time from report to settlement |
| Spikes without hiring | A storm triples the claims for a week; the queue does not triple | Size of the backlog |
| People do the judgement work | Handlers take the disputed and the large claims; the agent takes the routine ones | Share of staff time spent on exceptions |
| The same rules every time | Every claim is checked against the same policy wording, and every step is written down | Rule breaks found in audit |
| Lower cost for each piece of work | The cost of a settled claim drops, most of all when simple rules run before the model is called | Cost for each settled claim |
Worth it when the work is…
- high in volume and repetitive
- governed by clear rules
- easy to undo, or cheap to get wrong
- measurable: you can tell a good result from a bad one
Not worth it yet when the work is…
- rare, or different every time
- impossible to undo and high in stakes
- impossible to judge afterwards
- cheaper with a simple rule or a workflow
The catch
With nobody watching, there is no person to catch a mistake. Each problem below has an answer, and the answers together are what this course calls the harness.
| Problem | What it looks like | Where we meet it |
|---|---|---|
| It does not know when to stop | It loops all night, or says the job is done halfway | Stop 4: the limit in the loop |
| It can be tricked | A sentence hidden in a document becomes an order | Stop 5: the no-guardrails simulation |
| It can reach too much | It reads or changes things the person asking has no right to | Stop 5, second story |
| Small mistakes add up | A wrong reading at step 3 becomes a wrong payment at step 9 | Stop 6: reins that pull it back early |
| Nobody can see what it did | A customer complains and there is no record of why | Stop 8: the See family |
Next: enough talk. Stop 3 puts ordinary code and an agent side by side, doing the same job.