Skip to content

AI Agents

An agent is a planned, resourced AI worker. It is the agentic equivalent of a headcount line. Each agent has a name, a purpose, an accountable owner, and a model from the Models catalogue. You deploy it to projects, forecast its cost by runs, and reconcile the forecast against actual spend.

Agents live under Resourcing → Agentic workforce (/plan/main/resourcing/agents), on the Agents tab beside AI tools, Models and Rules. They do not sit on the Team Plan roster. They roll up in forecasts and on the roadmap alongside people, so a project's cost shows its people and its agents together.

This page covers the workforce view: how agents fit into the plan. The full guide lives in Agents.

What an Agent Record Represents

FieldDescription
NameDisplay label for the agent.
PurposeWhat the agent is for.
OwnerRequired. The accountable human. An agent cannot be saved without one.
KindOne of three kinds. See below.
PlatformThe provider or platform the agent runs on.
ModelA reference to a model in the Models catalogue.
Service accountLinked, or Not deployed. See Plan vs actual.

The three kinds:

KindWhat it is
Provider agentAn off-the-shelf agent from a provider's catalogue.
Built workflowA workflow you built on a provider's platform.
Custom workflowA bespoke workflow you build and run yourself.

An agent stores a reference to a catalogue model, never a copy of its price. When a catalogue price changes, every forecast that uses that model updates at once.

Copilots Are Not Workers

Copilots (Claude Code, Cursor, GitHub Copilot, Gemini) are tools people use. They do not get an agent record. Copilot spend loads onto the person who uses it, as part of that person's fully-loaded cost. Only autonomous agents get their own cost line.

This rule keeps totals honest. A person's copilot spend lands on the person once. An agent's spend is its own line once. Nothing is double-counted. The same product can be either: Claude Code run by an engineer is copilot usage on that engineer; Claude Code run as an autonomous workload is agent work. See the AI cost model for the full breakdown.

Deploying an Agent to a Project

You deploy an agent to a project with two settings:

  • Run cadence: expected runs per day, week or month.
  • Monthly spend cap: the most the deployment may spend in a month.

Flowstate forecasts the deployment as expected runs multiplied by per-run cost. You see a base case and a P90 estimate built from recent run volume, plus the headroom between the cap and the base-case forecast.

Deployed agents appear in the project's cost roll-up the same way allocated people do, so the roadmap and the cost forecast always show the combined cost of the humans and the agents on a project.

Plan vs Actual

Linking an agent to an AI service account is the deployment link. The service account is the agent's runtime identity: it has its own sessions and spend. Once linked, actual sessions and spend reconcile against the agent's forecast runs. You get plan vs actual for an agent, the same way you get planned vs actual cost for a person.

When you create a Custom workflow agent, Flowstate issues a one-time SDK key so your workload can report its usage back. The key is shown once. Store it securely before closing the dialog.

An agent with no linked service account shows "Not deployed" and reports no realised spend.

Where to Go Next

Flowstate Documentation