// agents

Hosted agents

An agent is a long-lived worker you configure once and reach from anywhere: a chat window, a Slack channel, a phone number, a schedule, a webhook, or another agent. It runs on our infrastructure, bills to your workspace, and keeps a full transcript of everything it did.

01// what a hosted agent is

One row, one runtime, many doors

A deployed agent carries a name and description, a system prompt, a model, a set of tools with a permission level on each, the channels it answers on, its schedules and webhook triggers, the knowledge collections it can search, and the workspace memory it writes to. Every execution of that agent, whoever or whatever started it, is a session.

Sessions are the unit of accounting and the unit of investigation. Cost is attributed per session, the transcript is stored per session, and a chat thread that continues over days is one session rather than many.

system promptmodeltools + permissionschannelsschedulesknowledgememory
02// draft to deployed

Nothing runs until you deploy it

An agent starts life as a draft. A draft exists in the workspace, shows up in the list, and can be edited freely, but it never receives input and it costs nothing. Deploying is the moment it becomes real.

  1. 01

    Give it a name and a prompt

    Deploy refuses an agent that is still called Untitled agent or that has an empty system prompt. Both checks run before anything is provisioned, so a refusal leaves no partial state.
  2. 02

    Clear the entitlement gate

    The organization needs an active feature plan, an active trial, or an active credit package. On the builder plans that is not enough on its own: the org must also hold the Agent Runtime module, which is what sells deploying there. Without either, deploy returns 402 DEPLOY_GATE_FAILED and the dashboard offers a card-backed trial.
  3. 03

    Deploy

    The runtime is provisioned, a per-agent gateway key is minted, and the agent flips live. From the first session it consumes workspace credits. Deploy is idempotent: calling it on a live agent returns the agent unchanged.

the gate is per organization

Trials are card up front and available once per organization, ever. The gate answers with the reason it let you through or turned you away — feature plan, trial, credit package, or none — so a failed deploy tells you which route to take.
03// where agents live

The index and the detail page

The agents index lists every agent in the workspace with search across names and descriptions, pinning, and teams. A team is a folder: agents can be moved into one, and deleting a team leaves its agents in place, just ungrouped. Archived agents collapse into their own section at the bottom, read only.

Opening an agent gives you six tabs: Overview, Sessions, Runs, Approvals, Analytics and Config. Talking to the agent is deliberately not one of them. Chat happens in the floating chat window, so the tabs stay an investigation and configuration surface rather than a conversation you can lose your place in.

04// every way an agent gets woken

Ten doors, one session model

Every surface that starts an agent stamps its own origin on the session it creates, so you can always tell what woke the agent. Conversations with a human deliver their reply back to the surface they came from. Background runs (schedules, webhooks, queues, inbox sweeps and agent-to-agent hand-offs) drain and bill but never post, and they stay out of the conversation list. Schedule and webhook fires additionally write a row on the Runs tab; the other background origins do not.

SurfaceWhat wakes it
Dashboard chatA person talking to the agent in the floating chat window.
SlackA message in the agent home channel, an @-mention in a mention channel, or a direct message.
iMessageA text from a confirmed phone binding to the shared agnt number.
ScheduleA cron-style trigger the agent owns. Produces a run row and a background work session.
Inbound webhookA third-party POST forwarded to the agent. Also a run row, also a work session.
Queue itemAn item drained off one of the workspace queues.
InboxEmail triage escalating a thread, or a reminder sweep coming due.
Another agentagnt_agents_dispatch handing work down. The sub-agent session records the caller.
The Agnt WizardThe builder opening a probe session against the agent while it configures it.
An MCP clientagnt_agents_session_start from a coding agent or any other MCP client.
05// managing agents from an mcp client

The lifecycle and invocation families

Everything the dashboard does to an agent is also a tool on the workspace MCP server, so a coding agent can build, deploy and drive agents without opening a browser. Both families are served by default; calls are metered against the Agent Runtime module on the builder plans.

no approval card here

In the dashboard the Agnt Wizard stages changes as a card a human clicks. These tools apply directly. Confirm with your user before calling deploy or archive.

Lifecycle

ToolWhat it does
agnt_agents_listEvery agent in the workspace with its draft, disabled and model state. Start here: every other call takes an agent_id from this list.
agnt_agents_getOne agent in full, including the config blob. Read it before patching, because the patch is a diff against exactly this config.
agnt_agents_createCreates a draft. A draft costs nothing and cannot receive input until it is deployed.
agnt_agents_update_configApplies a config diff immediately. There is no approval card on this surface, and a change to a deployed agent takes effect on its next session.
agnt_agents_deployA real deployment: it provisions the runtime, mints the gateway key and starts billing. Idempotent, so an already-deployed agent comes back unchanged.
agnt_agents_set_enabledStops or resumes an agent without tearing anything down. The reversible way to stop a misbehaving agent.
agnt_agents_archiveArchives the runtime and revokes the gateway key. Not reversible from this surface.

Invocation

ToolWhat it does
agnt_agents_dispatchFire and forget. Hands a task to another deployed agent and returns immediately; the run bills to the workspace under the target agent.
agnt_agents_session_startOpens a real, drivable session with a deployed agent. The synchronous counterpart to dispatch, and the right tool for a post-deploy smoke test.
agnt_agents_session_sendSends a follow-up message into that session. Each message runs the agent and bills compute.
agnt_agents_session_pollLong-polls the session status (idle, running, archived). Status only: reading the transcript is agnt_observability_read_session, in a separate family.

Reading what an agent actually said lives in the observability family, which is enabled independently. See sessions and observability.

06// limits and pricing

How many agents a plan carries

The agent cap is a plan limit, checked at deploy time. Drafts do not count against it, because a draft never runs.

PlanPriceAgentsSeats includedrpm
Solo$99/mo51120
Team$199/mo153300
Business$499/mo505600
Toolkit$19/monone until the Agent Runtime module is added1300
Toolkit All-Access$139/mo51600

On the builder plans, deploying is sold as the Agent Runtime module at $49/mo, with a 7-day free trial. It requires the Database & Canvas module ($19/mo), which the confirm flow adds alongside it. Holding it covers 5 agents, and further agents are $9 each.

published allowances

The quantity lines on the tiered modules (schedules, documents, pipeline items) are published allowances rather than enforced caps — what the tier is sized for, not a wall your agent hits mid-run. The agent count above is the exception: it is a real limit, checked before a deploy goes through.

Full plan and module detail lives on plans and Toolkit modules. Agent compute is paid from the AI wallet; data calls are paid from the data wallet. See credits and wallets.