What an AI agent build actually includes
An agent is not a single prompt. It needs four things to work at all: an orchestrator that decides the next step, a tool interface that defines what it is allowed to call, state so it remembers what happened earlier in the run, and a stopping condition so it hands back to a person or declares the task done. Get one of those wrong and the agent either does nothing useful or does something wrong with confidence.
We build all four against your actual systems, not a demo environment. That means the connector to your CRM, your EHR, your internal API with no public documentation, or the legacy database nobody outside your team fully understands. Your data does not leave your environment to make the agent work, and the code lands in your own repository and cloud account from week one.
- Agent orchestrator: plans the next action and holds state across steps
- Tool integrations: built against your real systems, not a generic connector list
- Eval suite: a golden set of inputs scored before every deploy, not read as five transcripts
- Guardrails: rules enforced in code, checked before an action executes
- Human-in-the-loop handoff: routes uncertain or high-risk calls to a person with context
- Audit log: every tool call and decision logged in your own infrastructure
Which workflows are worth an agent
An agent is the right shape for a workflow when three things are true at once: the task has multiple steps and at least one decision point, it calls tools or systems that already exist, and the cost of a wrong action is bounded and recoverable, or a human review step is cheap enough to insert before anything irreversible happens.
Some of what we have shipped: a voice-to-chart pipeline that drafts a structured clinical note from a dictated visit while a vision model reads an attached radiograph, both landing in the patient chart automatically. A fraud-detection engine that runs eight detectors over accounts-payable data and produces a per-jurisdiction PDF briefing with no analyst touching the pipeline. A compliance scanner that crawls a site, grades the privacy policy against a rubric, and returns specific fixes on a schedule.
An agent is the wrong shape when the task is genuinely single-turn, when nobody can define what "done correctly" means, or when a mistake would be expensive and unrecoverable with no guardrail able to catch it first. In those cases the honest recommendation is a well-scoped single-call AI feature, or a human process with AI assistance instead of AI autonomy, and we will say so in the first session rather than build an agent that looks good in a demo and fails in production.
How we keep agents from doing damage
Any vendor can demo an agent that does the right thing once, on camera, with a clean input. The question that separates a real build from a demo is narrower: how do you measure whether the agent did the right thing at scale, and what happens the moment it does not.
The eval is a fixed set of representative inputs with known-good outcomes, scored automatically before every deploy, so a change either makes the system measurably better or it gets blocked. The guardrail is a rule enforced in code, not a note in a prompt asking the model to be careful: never write to the billing system without a second confirmation, never message a customer outside business hours, never approve a transaction above a threshold without a human sign-off. When the agent's confidence is low or a guardrail trips, it hands off to a person with the context to decide fast, and every action it does take is logged in enough detail to audit and undo.
AI-assisted code goes through the same review gate as anything else we ship: a pull request in your repository, reviewed by the named engineer who owns it, typed contracts, and tests in CI. We do not train models on your client data.
What it costs and how the pod is staffed
Agent builds run as a pod, priced by capacity rather than by the hour. A single-track build, one workflow, its tool integrations, an eval and guardrail layer, typically fits a Builder Pod at $7,500 a month with a pod lead and a two-engineer bench. A build spanning two workflows, or a workflow plus a dashboard for a human to review edge cases, usually fits a Growth Pod at $10,000 a month with a three-engineer bench and bi-weekly strategy calls. Multi-department builds with their own compliance and access requirements move to a custom Enterprise Organization Pod.
Every tier is month-to-month with a 30-day cancellation notice by email. There is no per-hour billing and no change order when scope shifts, which it usually does once the first version of an agent meets real data. A small, single-workflow agent build runs 1 to 3 months end to end; a larger build spanning multiple workflows or heavier compliance requirements runs 3 to 12 months.