Per-seat desks (agents that draft, a person approves) and engines (standing agent crews Keystone operates, delivering counts on a cadence).
Claude models inside a deterministic harness: typed outputs, source citations on every draft, a golden-set evaluation before anything goes live.
Agents that send, submit, sign, or act on the world without a person. That is a design choice enforced by the architecture, not a policy.
Fixed price per desk from a two-week assessment. See what a custom AI agent costs.
What an AI agent is, and the question buyers ask
An agent is AI with a goal and tools. Given a job, it decides which tool to use next, reads what comes back, and keeps going until it has a result. That is different from a workflow, where the steps are written in advance. The agent vs workflow guide covers when each is right.
The question every buyer asks is "so these are AI agents?" The answer: the judgment parts are, yes. The desk is the harness they work inside, and that is what makes them safe to employ.
A desk is where agents report to work
A desk sits under one seat in your company: the estimator, the controller, the vendor coordinator. Five layers, one chassis.
- Connections. Read-only, into the systems the seat already uses: email, files, the project platform, the accounting system.
- Memory. An answer library that compounds with every job the desk touches. Half the value of the first build; the reason the second desk costs less.
- Plays. One per task: collect, extract, assemble, draft, flag. Written once, verified by a person, then trusted the way a machinist trusts a jig.
- Gates. Output is always a draft. A desk cannot send, submit, or sign. Enforced by the architecture.
- Triggers. Asked, then scheduled, then event-driven, as trust builds.
In production: a vendor onboarding desk that assembles packets from emails and PDFs and tracks insurance certificates, and a work order coordination desk that matches orders to crews and drafts dispatch notes, both at a Chicago-area construction and facilities contractor. Every draft cites its sources and every approval is logged.
Engines: agent crews Keystone operates
A desk responds; an engine runs. When work should arrive as a flow rather than a pile, the agents run in a fixed sequence on their own supply: read the sources, extract, score, rank, deliver. An engine never acts on the world. It writes to your record and hands a person a ranked list. It is Keystone-operated, so nobody on your team supervises it. Examples running today: a signal engine reading public records for funded construction work, and a trend engine scoring demand for a toy maker's design team.
How they are built
- Deterministic skeleton, small judgment organs. Routing, matching, formatting, and scheduling are plain code. The model is called only where a judgment is needed, with a typed schema for its answer.
- Claude models for the judgment calls, chosen for long-document reading and instruction following. Provider choice is not the point; the harness is.
- Provenance on every output. Each draft carries the sources it was built from, so approving takes seconds and auditing takes minutes.
- Single tenant. Each client gets a separate instance. Your code, your data, and a written offboarding process if you leave.
Evaluation before anything goes live
Every desk ships with a golden set: real past examples from the seat, graded by the person who sits in it, that the desk must pass before it handles live work and must keep passing when models change. Without evals an agent reaches a demo and never reaches production. With them, the coordinator can see the pass rate and decide how much to trust the queue.