Core concepts
Seven words that carry most of the product. Worth ten minutes before you build anything real.
Organization
The container for everything: agents, members, skills, connections, budget, and billing. Data is isolated between organizations, and your role is per organization — you can be an Admin in one and a User in another. Called a tenant in the API.
Agent
A configured worker: instructions, a model, tools, skills, memory, and files. It belongs to an owner and can be shared with colleagues. Agents don't share memory or files with each other, though they can talk to each other and search each other's memories.
Make a second agent rather than overloading one — an agent with a clear job and five tools works better than one with a vague job and thirty.
Task and session
A task is one unit of work. A session is the conversation it happens in; follow-ups continue the same session, so context carries.
Sessions persist and are searchable, and a running task survives you closing the browser.
Tool
A single capability — search the web, write a file, open a pull request. Some are built in; others need a connection first. Tools are attached per agent.
Skill
A reusable package of instructions and files that teaches an agent to do something a specific way. Where a tool is a capability, a skill is know-how. An agent picks the right one from its description — you ask for the work, not for the skill.
Routine
Work an agent does on a schedule, with nobody present — a morning digest, a poll, a reminder. You create one by asking the agent in plain language; there's no routines screen.
Budget
Caps on what agents can spend, in dollars. Hourly limits slow an agent down; daily and monthly limits stop it. Set organization-wide and per agent.
How they fit
Organization
├── Members (Users and Admins)
├── Connections GitHub, Slack, … — set up once, org-wide
├── Skills reusable know-how, governed centrally
├── Budget hourly / daily / monthly caps
└── Agents
├── owner + people it's shared with (Viewer / Editor)
├── instructions and model
├── tools some backed by a connection
├── skills attached from the org's set
├── memory durable, per agent
├── files workspace / attachments / skills
├── routines scheduled work
└── sessions conversations, searchable
Two things people take longest to internalise:
- Connecting a service isn't the same as giving an agent access to it. Connecting is organization-wide; attaching the tool is per agent.
- Organization role and agent access are different axes. Being an Admin gives you every agent; owning an agent gives you one.