Service

AI agent development

Products where an LLM does the work — assistants, copilots, and automations with real guardrails, evals, and cost controls.

AI agent development is building a product where a language model does the work: reads the email, drafts the reply, fills the form, answers from your documents, chains the tools. The hard part is not getting a model to do something impressive once — anyone can do that in an afternoon. The hard part is the plumbing that makes it dependable enough to charge for: tool design, guardrails on what the agent may and may not do, evals that tell you when quality slips, and cost controls so a busy day doesn't produce a shocking API bill.

That plumbing is what an engagement with us buys. We design the agent's tools and prompts, wire it to the systems it works across, and put an eval harness around it from the first week — a set of graded test cases that turns 'it seems better' into a number. Guardrails are explicit: what the agent can touch, what requires a human, what gets logged. The interface on top is deliberately plain, because the agent is the product.

We build agents AI-first, which sounds recursive but matters for one practical reason: iteration speed. Your first prompt architecture will not survive contact with real usage, and neither will your second. Because the surrounding code is cheap for us to change, re-architecting the agent's brain in week four is a normal week, not a renegotiation. The estimator prices the typical agent build in two minutes; the cost pages under this service show that number against what senior AI engineering time costs in your city.

Costs at runtime stay yours and visible: the agent runs on your API keys, under usage caps we set together, with per-request cost logging from day one. When the engagement ends you own a working agent, its eval harness, and the numbers that tell you whether the next prompt change made it better or worse.

What it costs

A typical scope prices at $9,500–$14,000 built AI-first and ships in 4–6 weeks — here is the same scope under all three approaches, from the estimator's public constants.

Vibecoding

$9,500–$14,000

4–6 weeks

Hybrid

$13,000–$19,000

6–8 weeks

Traditional

$17,500–$25,500

9–11 weeks

Lean

A working agent on one workflow with a minimal interface — the fastest way to find out if the model can do the job.

$6,000–$8,500

Standard

The typical build: the agent, accounts, the integrations it works across, and an admin view with cost logging.

$9,500–$14,000

Ambitious

Adds analytics, file handling, and a month of post-launch tuning — for agents going in front of customers.

$13,500–$19,000

Price your configuration

What's in the price — and what isn't

Included in every engagement

Agent and tool design

The prompts, tool definitions, and orchestration that make the model do the job — designed for your workflow, not adapted from a demo.
Guardrails and permissions

Explicit rules for what the agent may do alone, what needs a human, and what is off-limits — enforced in code, not in the prompt.
An eval harness

Graded test cases run on every change, so quality is a number you watch instead of a vibe you argue about.
Cost controls and logging

Per-request cost tracking and usage caps on your API keys from the first deploy.
A plain interface on top

The screens needed to use, supervise, and correct the agent — deliberately simple, because the agent is the product.
Your keys, your accounts, your code

Runs on your model API accounts and deploys from your repository. Nothing routes through us in production.

Explicitly not included

Model API usage fees

The agent runs on your OpenAI, Anthropic, or other API accounts; usage bills you directly, with caps and logging so it stays predictable.
A guarantee the model is never wrong

Language models make mistakes. What we build is the harness that catches, measures, and contains them — anyone promising zero errors is selling something else.
Training or fine-tuning custom models

Agents here run on hosted frontier models. Fine-tuning is its own project with its own economics, scoped separately if evals show you need it.
Ongoing prompt maintenance

Models and providers change; the support retainer in the estimator covers post-launch adjustment, or your team takes it over with the eval harness.
Compliance certification

Sane defaults on data handling are included; SOC 2 or HIPAA audit work is a separate project.

AI agent development cost, city by city

The price is the same everywhere we work — these guides put it next to each city's own market rates, agency floors, and hiring costs, with sources.

Common questions

Which models do you build on?

Hosted frontier models — Anthropic and OpenAI, chosen per task in evals rather than by loyalty. The agent runs on your API accounts either way, so switching models later is a configuration change, not a rebuild.

How do you keep API costs predictable?

Usage caps and per-request cost logging are wired in from the first deploy, on your own keys. You see cost per task in the admin view, and the caps mean a busy day degrades gracefully instead of billing surprisingly.

What stops the agent from doing something it shouldn't?

Guardrails enforced in code: an explicit list of tools it can use, actions that require human approval, and hard limits it cannot cross no matter what the prompt says. The model is never the only thing standing between the agent and a mistake.

How do I know the agent is actually good?

An eval harness — graded test cases from your real workflow, run on every change. It ships with the product, so 'did the last change help' stays a measurable question after we're gone.

Do I need my own OpenAI or Anthropic account?

Yes, and that's deliberate: your keys, your data agreements, your usage bill, your ownership. We set the accounts up with you at kickoff — it takes minutes and means nothing about your product routes through us in production.

Why are agent MVPs cheaper than SaaS MVPs in the estimator?

Smaller surface area: an agent MVP is one deep workflow with a plain interface, while a SaaS MVP usually carries more screens, roles, and billing states. The base prices in the estimator reflect that — and grow the same way if you add modules.