How Auto Lab works
The ideas behind Auto Lab in one page: goals, Chat, task agents, plans, approvals, the Brain and learning.
This page explains the parts of Auto Lab and how they fit together. Read it once and the rest of the docs will make sense in any order.
Organizations and goals
Your organization holds your goals, the people who work on them and shared settings such as members, roles and single sign-on. You can belong to more than one organization.
A goal is one outcome you want Auto Lab to work toward, such as "follow up with every new customer in their first two weeks". A goal has:
- one agent that owns the work, plus task agents it hands pieces to
- a plan of tasks for the current week
- one running conversation, Chat
- a Brain: what its agents know and the rules they follow
- the connectors, channels and secrets it may use
You can run many goals side by side. The Goals home shows every goal you can see, grouped by whether it needs you, is in progress or is paused.
Everything a goal knows is kept in a repository
Each goal has its own private git repository. Auto Lab creates and manages it for you, or you can connect one on GitHub. The goal's instructions, memory, skills, plan and settings are files in that repository.
This is why every change to a goal has an author, a date and a history, and why a change can be reviewed before it lands or undone after. You never need to open the repository to use Auto Lab. Developers can clone it; see The goal's repository.
The agent, task agents and runs
The goal's agent runs on Auto Lab's servers. It reads messages, keeps the plan up to date, decides what to do next and talks to you. When a piece of work needs focus, it hands that piece to a task agent with one job:
| Task agent | Does |
|---|---|
| Research | Answers one question from the web, the goal's memory and its connected apps |
| Docs | Drafts or updates one document or message |
| Mail reader | Turns an incoming email into a short summary. It has no tools, so it cannot act on what the email says |
| Browser | Uses websites, including ones that need a login |
| Coder | Changes files in a repository and hands the result back as a change request for review |
Each hand-off is a run. Runs show up as cards in Chat and on the goal's Runs page, where you can follow them, including a live view of the browser.
When work needs a real computer, for example to run a script or edit files, the agent starts an isolated cloud machine for it. A Coder task or a scheduled trigger gets a session: a machine with its own branch of the goal's repository, whose work ends in a change request. Machines stop when the work is done or has been idle for a while.
Chat: one conversation per goal
Chat is the goal's single running thread. Messages from the web app and Slack land in it, and so do Microsoft Teams, iMessage and email when they are switched on for the goal. The agent has the whole story wherever you write from. When it asks you something, you can answer from any of those places, and its reply goes back where you asked.
Chat also shows what happened in between: approvals waiting for you, questions, task cards, and short notes such as "Learned: …" with an Undo button.
Plans, tasks and schedules
At launch you approve a plan for the first week. Each task is owned by the agent or by you. On the Overview, tasks sit under In progress, Blocked, Planned and Done this week as the work happens.
Nothing in Auto Lab runs all the time. The agent wakes up when:
- a schedule fires: the routine, a check-up it set for itself, or a reminder
- a message arrives in Chat or a channel
- something it is waiting for happens, such as a reply to an email or a request from another app (a webhook)
The agent re-plans as it learns more. How much of that it may do without asking depends on the goal's autonomy. See Plans, schedules and triggers.
Autonomy, approvals and Needs you
Each goal has a Delegation level, set during setup as On its own:
| Level | Meaning |
|---|---|
| Asks before acting | Every action in a connected app, even reading, waits for your approval |
| Acts within your rules | Your rules decide. By default, reading runs and anything that writes or deletes asks first |
| Acts on its own | Actions run without asking, unless a rule says otherwise or the connector is set to Ask before every use |
Approvals, reviews, plan changes and tasks assigned to you collect in Needs you on the goal's Overview. Questions from the agent wait above the Chat composer. When you are away, both reach you in the channel you picked during setup. When you approve an action, you can also select Approve and make it a rule so the same kind of action runs without asking next time.
Bigger changes arrive as a change request: a proposed set of edits to the goal's repository that you can review, approve or send back. A plan revision, a new skill or a code change can each be a change request. An agent never approves its own. See Approvals and autonomy.
The Brain and learning
The Brain is where you see and edit what the goal's agents know: agents, skills, memory, tools, connectors, schedules and more. Each item shows who changed it last and where it is used.
The goal learns from its own work. After a stretch of turns, and again overnight, it reviews what happened. Small, safe lessons, such as a preference you stated or a better way to do a task, apply at once and appear in Brain › Learning with Undo. Larger changes, such as rewriting memory or editing a skill a person wrote, wait for your approval. See Learning.
Keeping credentials safe
- Connectors keep their credentials on Auto Lab's servers. The agent asks for an action; the server makes the call.
- When the agent needs a website login, it sends you a login link. You type the password on that page, never in Chat. See Website logins and secure links.
- Secrets you add can be granted to specific agents only. See Secrets.
Words used in these docs
| Term | Meaning |
|---|---|
| Organization | The group that holds goals, members and shared settings |
| Goal | One outcome Auto Lab works toward, with its own agent, plan, Chat and Brain |
| Chat | The goal's one running conversation, shared across web and channels |
| Task agent, run | A focused helper the agent hands work to, and one piece of work it did |
| Session | An isolated cloud machine with its own branch, used for code changes and scheduled triggers |
| Needs you | The list on a goal's Overview of approvals, reviews and tasks waiting on a person |
| Change request | A proposed change to the goal's repository that a person reviews |
| Brain | Everything the goal's agents know, and its history |
| Connector | Access to an outside app or API |
| Channel | A place the agent talks to people: Slack, Teams, iMessage or email |