The demo took a weekend. The production version has been “two weeks away” for a quarter. Between them sits work nobody demoed: authentication into six internal systems, a queue that survives a deploy, an approval screen for the cases the agent should not decide alone, and a test suite that catches the regression before your customer does.
An AI agent developer is the person who does that work.
What gets built
The agent
Planning and tool-calling loop, scoped to a task small enough to verify. Budgets, retries, cutoffs, and an escalation path to a human when confidence or permissions run out.
The tools
Each integration wrapped in a narrow, validated, idempotent contract. Real auth, real rate limits, real pagination, real error handling. This is where agents mostly fail and where most of the engineering time goes.
The runtime
Queues, workers, durable execution for long-running tasks, and a deploy that does not lose in-flight work. Temporal or a job runner when the workflow must survive a restart; something far simpler when it does not.
The supervision surface
Approval queues, trace viewers, run history, cost per task, and a kill switch that any ops lead can reach without paging an engineer.
The tests
Golden tasks in CI, adversarial cases alongside the happy path, and a regression gate that blocks a merge when quality drops.
Where agents pay off first
Teams get the fastest return where the work is high-volume, rule-shaped, and currently done by a human copying between two systems:
- Support triage and drafting - classify, retrieve policy, draft a reply, route the ones that need judgement.
- Document processing - extraction from contracts, invoices, claims and filings, with a confidence threshold routing the ambiguous cases to a person.
- Internal research - pull from your own corpus, cite sources, and refuse to answer when the corpus does not cover it.
- Operations automation - reconciliation, data hygiene, ticket enrichment, onboarding and offboarding runbooks.
- Engineering workflows - review triage, migration assistance, test generation, dependency and release hygiene.
And where they do not pay off: anything that needs a guaranteed-correct answer every time with no human in the loop, anything where the cost of a wrong action is unbounded, and anything you could solve with a SQL query and a cron job. You will get that answer on the first call, for free.
Working style
Small scope, shipped behind a flag, measured before it is widened. Your engineers in the pull requests from week one. Written decisions, so the reasoning survives after the engagement ends. No runtime to license, no platform lock-in, and the code, prompts and evals are yours from the first commit.