AI agent development, for agents that act on real systems
An agent that answers questions is a chatbot. An agent that books, updates, refunds, escalates and files is a different kind of system, because every one of those actions can be wrong, and some cannot be undone. We build the second kind. That means the boring parts get most of the attention: what the agent is allowed to do, which actions stop for a human, how it behaves when a tool fails, and how you find out when it drifts. The model is the easy part.
Who this is for
- - A support or operations team that wants an agent to resolve tickets end to end, inside the systems you already run, not just draft replies.
- - A SaaS company adding an agent to its own product and needing it to act on customer data with the customer’s permissions, not a blanket service account.
- - A team that tried an agent platform and hit the wall: logic the platform cannot express, or write access to a system of record it cannot reach.
- - A business in a regulated space where every action needs a log, a limit and a person who can be named as responsible.
What you get
- - A written action inventory before any code: what the agent may do, what needs approval, what it may never do.
- - Tool integrations into your systems of record with least privilege access, inheriting the user’s permissions where one exists.
- - Approval gates in front of the actions that cannot be undone, and nowhere else, so the person approving actually reads them.
- - An evaluation set of real scenarios, including the adversarial ones, run before every change.
- - Observability per step: which tool was called, with what, and what came back, so a wrong outcome can be traced.
What it costs
There is no honest fixed price before discovery, and anyone who gives you one is guessing. What we can say in writing after two weeks: the build cost, the running cost per month at your expected volume, and what moves each. The build is driven by how many systems the work must touch and how strict the accuracy, latency and compliance bars are. The running cost is model calls plus infrastructure, and it can differ by ten times depending on the choices made in week one. How we price AI work.
How long it takes
Discovery, two weeks: we map the problem, the systems involved and the success criteria, and put the numbers in writing. Pilot, four to six weeks: one capability, live, on a slice of real traffic, measured against those criteria. Then production hardening and scale out. Each stage is a separate agreement with no minimum commitment, so you can stop after discovery with a plan you own. How we work.
What you own afterwards
Everything. The code, the prompts, the evaluation set, the infrastructure configuration and the documentation, in your accounts, under your name. We do not hold anything hostage and we do not license our own platform to you. The evaluation set matters most: it is the thing that lets your team, or any other team, change the system later without breaking it. The deliverables clause.
Questions buyers ask
What does an AI agent development company actually deliver?
A working agent connected to your systems, plus the parts a demo skips: a written inventory of what it may do, approval gates in front of irreversible actions, an evaluation set of real and adversarial scenarios, and per step logging so any wrong outcome can be traced. All of it in your accounts.
How much does custom AI agent development cost?
It is driven by how many systems the agent must act on and how many of those actions are irreversible, because each of those needs integration work, a permission model and a test. After two weeks of discovery we put the build and the monthly running cost in writing. A single well bounded agent is a very different project from an agent that touches six systems.
Should we build a custom agent or use an agent platform?
Use a platform when the task fits its shape, volume is modest and the agent only needs read access. Build when the agent must write to systems the platform cannot reach, when the logic does not fit the platform’s template, or when you must own the result. Our build versus buy article works through the numbers at published platform rates.
How do you keep an AI agent safe?
By deciding, before any code, which actions are consequential and gating only those; by giving the agent least privilege access that inherits the user’s permissions; by testing against adversarial inputs including prompt injection; and by logging every tool call so drift is caught early. Gating everything is close to gating nothing, because a control that fires constantly stops being read.
Before you decide
If you want the technical background first, read how AI agents work. Then, from our engineers:
Tell us what the agent needs to do
Two weeks of discovery puts the cost, the timeline and the plan in writing before you commit to anything. Or start with the free audit.