Service — from $25k
Agents that are allowed to act, because nothing irreversible happens unsupervised.
The demo is not the hard part. Any competent engineer can get a model to call a tool. The hard part arrives the first time it is confidently wrong while holding your send button, your database, or your payment API — and it will be, because that is what the technology is.
Most agent projects answer that with a line in the system prompt. That is not a permission system. It is a request, made to the one component in the stack that is known to occasionally ignore requests.
So the split is drawn in the runtime instead. Reading, drafting, summarising and routing run unsupervised, because the worst case there is a wasted minute. Everything that cannot be undone stops and waits for a person. That single boundary is most of the difference between an agent you can put in production and a demo you keep having to supervise.
The short answer
AI agent development from Code and Proof starts at $25,000 and builds agents that are allowed to act in production. Nothing irreversible — sending, charging, deleting, publishing, filing — happens without a human approving it first. Side effects carry idempotency keys so a retry is a no-op rather than a second invoice, tool access is enforced in code rather than requested in a prompt, and per-run cost ceilings stop a loop becoming a billing incident.
What gets built
Six parts. The model is one of them, and it is not the one that takes the time.
The approval gate
The agent proposes; a person disposes. The hard part is not the button — it is holding the run's state coherently while it waits, sometimes for days, and resuming correctly when the answer finally arrives.
Idempotent side effects
Every agent runs twice eventually: a retry, a redeploy, a duplicate webhook. Every effect carries a key so the second run is a no-op rather than a second invoice sent to a real customer.
Tool boundaries in code
What the agent may touch, enforced by the runtime rather than requested in the system prompt. A prompt is not a permission system. It is a polite request to a component that is known to be occasionally wrong.
Cost ceilings
An agent in a loop is a billing incident. Per-run and per-day ceilings, enforced in the runtime, with a defined behaviour on hitting them that is something other than “crash”.
A readable transcript
What it decided, what it called, what came back, what it did next — as prose a person can read at 2am, not as a dashboard that summarises away the one line that matters.
Self-healing
Queues stall, machines sleep, APIs rate-limit. An agent expected to run unattended for months has to notice its own gaps and reassign the work it missed, rather than going quiet and looking fine.
Where the gate goes
The rule is reversibility, not risk appetite. If an action cannot be undone by the system that took it, a person approves it first. This is the table that gets filled in on day one of an engagement, because arguing about it later means arguing about it in production.
This is not theoretical
Three agent systems on this site already work the way described above, and each one is written up in enough detail to argue with.
What it costs, and why
Agent builds start at $25,000. That is the agentic band in this site’s own published cost guide, and it is deliberately not undercut here — a number on a sales page that contradicts the research on the same domain is worth less than no number at all.
An agent costs more than a pipeline for a reason that has nothing to do with the model. The expensive parts are the approval gate holding state across days, the idempotency keys, the ceilings, and the transcript. All of them exist for the same purpose: to make being wrong survivable. Skip them and the build is cheaper and the operation is not.
If the agent is already live and misbehaving, this is the wrong page. Start with the $2,500 reliability audit so the repair is scoped from evidence rather than an estimate.
Common questions
- What does it cost to build an AI agent?
- AI agent development from Code and Proof starts at $25,000, which is the agentic band in this site's own published cost guide. An agent is more expensive than a pipeline because the expensive part is not the model — it is the approval gate, the idempotency keys, the cost ceilings and the transcript, all of which exist to make being wrong survivable.
- What does an AI agent need before it can act in production?
- A human gate, enforced in the runtime rather than requested in a prompt. Sending, charging, deleting, publishing and filing all wait for a person to approve them. Every side effect carries an idempotency key so a retry is a no-op rather than a second invoice, and a per-run cost ceiling stops a loop becoming a billing incident.
Related
and I will tell you where the gate goes.