Agent engineer
The one who makes a self-deciding loop hold together
An agent engineer builds systems handed a whole task rather than a question, and whose failures have to be planned for, like a stage manager preparing the recovery for the night an actor goes off script.
- Floor 3The Agent
- P2Writes the application
- Emerging titleIt appeared recently. What it covers is not yet stable from one employer to the next.
What they answer for
The deliverableA loop that carries a task through unsupervised, with its trace, its stopping points and what it is allowed to do. If they leave, nobody can say why it stops.
The profile
Five activities, scored 0 to 3- ResearchProducing knowledge that does not yet exist.
- BuildShipping a system that runs, deploys and breaks.
- OperateKeeping it in production: cost, incidents, drift, on-call.
- VerifyMeasuring, testing, attacking. Producing a verdict that holds.
- LeadDeciding, persuading, driving adoption, answering to others.
What the work is
What changes
The difference from the floor below fits in one sentence: you no longer ask for an answer, you hand over a task. The system then picks its steps, calls tools, gets things wrong, starts again. What was a call becomes a journey, and a journey needs bounds.
The real work
The real work is not making the loop run, which takes a few dozen lines. It is deciding what happens when it fails. Where it stops, what it may do without asking, what it must leave behind so that anyone can understand afterwards.
The trap
The trap is the demonstration. A loop that completes a task in front of witnesses proves nothing: the question is what it does the time it fails. An advert that talks about capabilities and never about guard rails is describing a prototype.
A week in the role
- Loop code: tools, stopping points, retries, a budget of steps.
- Reading traces, the only way to understand a decision you did not write.
- Guard rails to place: what the agent does alone, what needs a human to agree.
- End-to-end measurement, the task being done or not done, never half.
Ways in
- From application engineering around a model, which is the ordinary way in.
- From distributed systems, where being used to retries and partial state counts for more than knowing the model.
- What the role does not require, whatever the advert says: training a model.
Reading an advert
2 signs- The advert describes capabilities and never a guard rail.
- That is a prototype role. The trade begins at the question of what happens when the loop gets it wrong.
- The word "autonomous" keeps coming back with no limit named anywhere.
- Unbounded autonomy is not a technical goal, it is a decision nobody has taken. Somebody will have to, and it will be you.