Agents

An agent is a planned, resourced AI worker. It has a name, a purpose, an accountable owner, and a model from the Models catalogue. It is the agentic analogue of a headcount line: you forecast it by expected runs and per-run cost, the same way you forecast a person by salary and allocation.

The list lives at Resourcing → Agentic workforce → Agents (/plan/main/resourcing/agents).

The three kinds

Every agent is one of three kinds. You pick the kind first when creating one.

Kind	What it is
Provider agent	An off-the-shelf agent from a provider's catalogue.
Built workflow	A workflow you've built on a provider's platform.
Custom workflow	A bespoke workflow you build and run yourself.

The kind decides what you fill in next. A provider agent is picked from the provider's catalogue and can carry the agent's ID on that provider. A built workflow names the platform it runs on. A custom workflow is yours end to end, so Flowstate gives it a reporting key instead (see below).

Creating an agent

Click Add agent. The drawer asks for the kind first, then the details:

• Name. For a provider agent, the name defaults to the provider's agent name.
• Purpose. What the agent is for. One sentence is enough.
• Provider or platform. Where the agent runs, depending on kind.
• Owner. Required. An agent needs an owner before it can be saved. The owner is the person accountable for what the agent does and what it spends.
• Model. Picked from the Models catalogue. The agent stores a reference to the model, so catalogue price changes flow through automatically.

Custom workflows get an SDK key

Creating a Custom workflow issues a one-time Flowstate SDK key. Add it to @flowstate/ai-sdk-node so your workload reports its usage back to Flowstate.

WARNING

The key is shown once. Copy it before closing the drawer. Lose it and you have to reissue.

Provider agents and built workflows don't need an SDK key. Their usage arrives through the provider connection or the linked AI service account.

The list view

Each row is one agent:

Column	What it shows
Type	Provider agent, Built workflow, or Custom workflow
Owner	The accountable person
Model	The catalogue model the agent runs
Platform	The provider or platform it runs on
Cost	Realised AI spend over the selected date window

Cost is realised spend, not the forecast. If an agent's model has no catalogue price yet, the cost cell shows Awaiting price data until the price lands.

Deploying to projects

Agents don't get dragged onto a schedule like people. You deploy them, deliberately and with a cap. From a project, open the Deploy agent drawer:

1. Pick an agent from the pool. Only agents linked to a deployed service account appear in the pool. Copilots never do.
2. Check the forecast. The drawer shows a base-case monthly forecast and a P90 figure. The P90 comes from recent weekly run volume, so it needs at least two weeks of run history before it diverges from the base case.
3. Set a monthly spend cap. Optional but recommended. Deployments without a cap show as Uncapped.
4. Check the headroom. Headroom is the cap minus the base-case forecast. Negative headroom means the cap is below what the agent is already on track to spend.

The drawer also shows a recommended config: the cheapest model tier and tool scope that clears the agent's observed per-run demand.

Each deployment carries an expected run cadence, expressed as runs per day, week, or month. You can see the cadence and cap for every deployment on the agent's Schedule tab.

The deployment link

Linking an agent to its AI service account (Settings → AI → AI Service Accounts) is the deployment link. The service account is the production identity that actually calls AI providers. Once linked:

• The service account's actual sessions and spend reconcile against the agent's forecast runs. Plan versus actual, exactly like a person's planned versus actual cost.
• The agent's realised-spend figures fill in. An unlinked agent shows "No deployment linked yet."
• The agent becomes eligible for the project deploy pool.

See Spend reconciliation for how actuals land.

The agent record

Click a row to open the agent. Five tabs:

Tab	What's there
Overview	Name, purpose, owner, kind, platform, model, region, and the service account link status (Linked or Not deployed). A header stat shows realised spend for the last 30 days.
Schedule	Project deployments, each with its run cadence and spend cap.
Activity	What the agent did in the last 30 days: projects worked on, run counts, and any unattributed work not linked to a project.
Changes	The change history for the record.
Comments	Discussion on the agent.

How agents roll up

Agents sit in your plan next to people:

• Forecasts. Agent cost lines appear in the cost forecast alongside salaries and contractor rates.
• Roadmap. Deployed agents show on project schedules with their cadence and cap, where a person would show FTE.
• Caps and alerts. AI budget caps live at Finance → Budgets → AI caps. When an agent's spend breaches its cap, an alert fires on Insights → Agent insights → Alerts.

For how agent spend relates to copilot spend and people costs, read The AI cost model.