Nexto Labs · agents for the work nobody wants

Hand over the job, not the tool.

One agent, one recurring job, owned end to end: read the queue, decide, act, escalate when unsure. For ops and engineering leads at 50–500 person companies.

See one agent, in full Six stages · one job · a baseline it has to beat

Is your job agent-shaped?

Four things have to be true at once. Most jobs fail on the third, and they fail quietly. The work looks routine until you write down what a correct outcome is.

Read them against a real job. If two are false, we say so on the call.

High volume: Dozens of times a week, not four times a month.
Judgment-light: Two experienced people would agree on the outcome.
A clear right answer: Gradeable afterwards, without a meeting.
Reading and typing: Done today by looking things up and writing them down.

One agent, in full: inbound support triage.

No client name on this one. We’re new. The mechanism is the proof.

The job. Seven of every eight new tickets are routine: where’s my order, change the address, refund this. Reading and typing. The eighth needs a person, who spends the day on the other seven. The same queue sits in an online shop, a clinic, a wholesaler, a broker’s office. Nobody there writes code.

How it’s checked. The eval set exists before the agent does: real tickets, each with the outcome your team agreed was correct. A build that scores lower doesn’t ship.

When it’s wrong. Wrong intent goes to a human, and that ticket joins the eval set the same day. Irreversible actions are proposals a person approves. Uncited answers are blocked.

What it’s given, and where it stops
Ticket queue	Read, search, tag, assign.	Cannot delete or merge.
Order lookup	Read-only.	No writes, ever.
Refund API	Drafts a refund.	A person approves every one.
Shipping address	One write, pre-dispatch.	Blocked after dispatch.
Knowledge base	Reads and cites.	An uncited answer is blocked.
Reply	Sends on the seven known intents.	The eighth goes to a human.

Not settings. Boundaries the agent has no tool to cross.

Inbound support triageClassify the queue, answer the known, escalate the rest.
Prospect researchA list of companies in, the same twelve fields out, with sources.
Document intakeStructured fields out of PDFs that arrive in forty layouts.
Internal question answeringAnswers from the wiki, cited, or “I don’t know”.

The jobs we take on.

Four archetypes, because in each one “done correctly” can be written down. If yours isn’t listed but passes the four tests, it still qualifies.

How an agent gets built.

Six stages, in order. Two and three are the ones everyone wants to skip, and the ones that decide whether anyone can tell if this worked.

The sheet · four fields, filled in by the person who does the job today

Brief it like a new hire, including what doing it badly looks like.

Stage 1 / 6

What it costs.

Three shapes, in order. Each one is scoped and quoted before you commit. Stopping after the first is a fair outcome.

Engagement	What you get
Job spec + eval set2–3 weeks	The spec, the measured baseline, and a graded eval set, all yours to keep. If the job isn’t agent-shaped, you find out here.
Build + shadow run6–10 weeks	The agent, its tools, its guardrails, and a shadow run. It graduates only if it beats the person doing the job today.
OperateMonthly	We run it, watch for drift, and keep the eval set current. Or we hand all three over and your team does.

Tokens and infrastructure are billed to you by your own providers. We don’t resell either.

When is an agent the wrong tool?

Low volume, contested answers, inputs nobody wrote down, or a mistake you can’t walk back. A job that runs four times a month never pays back its eval set. We’d rather lose the deal here than in month three.

What happens the first time it’s wrong?

You hear it from monitoring, not from a customer. It ran in shadow first, and irreversible actions are proposals a person approves. The case joins the eval set the same day.

Who owns the code and the eval set?

You do. All three, in your repository, from the first commit. The eval set matters most: it keeps its value if you replace the model, the framework, or us.

Whose cloud does it run in?

Yours. Only the model call leaves your perimeter, under your provider’s terms, not used for training. Data that must never leave, the agent never reads.

What does it cost to run per month?

Model tokens plus hosting, billed to you by your own providers. We don’t resell either. We estimate it from your baseline before you commit to the build.

How much of our team’s time?

A few hours a week at the start, from the person who does the job today. Nobody else can supply the correct answers. After that, a review cadence.

The awkward questions.

Answered here, so you don’t have to ask. The first matters most.

Three artefacts, and you own all three.

They sit in your repository from the first commit. They make the second agent cheaper than the first, and they hold their value if you replace the model, the framework, or us.

The job spec: The job briefed like a new hire, including what doing it badly looks like.
The measured baseline: Volume, hours, cost, error rate. The numbers the agent has to beat.
The eval set: Real cases with known-correct outcomes. The one artefact that survives a change of model, framework, or vendor.

jose@nxtolabs.com