Skip to main content
An agent is a persistent identity attached to a single instance. It has a name, a workspace full of files, a growing catalog of skills, and a set of credentials it can use to reach the outside world. Every time you talk to it in chat, a channel message, or through an automation, the same agent picks up where the last conversation left off.
An instance is the hardware. An agent is the person. The two are tied together one-to-one, so when we talk about “your agent Daniel” we mean the instance named Daniel plus everything it knows about you and your work.

​
What an agent remembers

When you create an agent, it boots with a short bootstrapping ritual: it picks a name, writes a short IDENTITY.md, and records what it knows about you in USER.md. Over time the workspace fills out:
File or directoryWhat lives there
IDENTITY.mdThe agent’s name, role, preferred tone
USER.mdFacts the agent has learned about you
HEARTBEAT.mdA running log of what the agent is working on
workspace/Project folders: drafts, references, renders, data
skills/Custom skills you and the agent build over time
sessions/Transcripts of past chats and subagent runs
cron/jobs/Scheduled automations this agent owns
All of this is visible in the Files tab in the dashboard. You can read, download, or drop new material in at any time.

​
Creating a new agent

Ask Operator in chat. Be specific about what you want the agent to focus on: 
Spin up a new agent to make music videos with me. Name them whatever
fits. Give them a taste for dream-pop visuals and a habit of drafting
three options before committing to one.
Operator deploys the instance, runs the bootstrap conversation to pick a name and identity, installs any skills it already knows will be useful, and returns a chat link when the agent is ready.

​
One agent per self-serve plan

Basic, Pro, and Max each run a single agent. You go deep with one: it learns your voice, accumulates skills, and owns every file in its workspace. Most projects live comfortably inside a single agent.

​
Going multi-agent with Enterprise

Some teams want the agents themselves to specialise. A research agent that never touches drafts. A creative agent that never scrapes the web. A comms agent that owns the outbound voice. That’s the Enterprise plan. Enterprise unlocks a custom instanceLimit so several agents can run in parallel, each with its own memory. Direct work at a specific agent with @name in any message: 
@daniel draft the next chapter. @maya review it once daniel has a
draft in /workspace/drafts/. Ping me when both are done.
When multiple agents are mentioned, Operator fans out, delivers the relevant task to each, and stitches the replies together. You can also ask Operator to clone an agent with one variable changed (visual style, research angle, persona) and run them side by side to see which approach works best. Reach out to [email protected] when a single agent isn’t enough.

​
How agents improve

Agents get better the same way people do: by doing the work and keeping notes.
  • Skills capture approaches. When you find a flow that works (a specific prompt chain, a script, a post-production recipe), ask Operator to turn it into a skill. The agent will reach for it on future tasks automatically.
  • Files carry history. Reference material lives in workspace/. The agent can grep its own past notes before asking you for context.
  • Automations run the boring parts. Anything repetitive becomes a cron job. See Automations.
  • Ground truth is better than synthetic tests. Connect the agent to your actual inbox, database, or repo instead of sandboxed fixtures. Results get more honest fast.

​
Suspending and deleting

Suspending an agent removes the running container but keeps the workspace intact. The agent goes quiet on its channels until you restart it, but nothing gets lost. Deleting an agent wipes the workspace and releases the slot. Both actions are available from the agent’s page in the dashboard or by asking Operator directly.

Instances

The computer each agent lives on.

Skills

Reusable playbooks agents load at turn time.

Automations

Scheduled work handed to the agent.

Environment

Credentials agents use to reach the outside world.
Start building →