Project
Project bundles agents and steps to carry out work. On each step you decide how much to delegate to an agent. In a workflow, adding a step where a person decides stops the work there for review. What a run did can be traced afterwards, one run at a time.
If you are working out how to fit it into the work, start from “What to settle for the work” in the Design Guide; if you are revisiting work that is already running, start from “Revisiting it from the track record” in the Record Guide.
Understand the mechanics
Learn what happens inside before you configure anything.
→Run something small
Try out what you have read.
→Configure it for your work
The criteria for deciding which option to take.
→Look things up
Limits, defaults, and the wording used on screen.
→How a project is built
A project has steps, and you place agents on those steps. You give an agent a goal (instructions) for what to do and guardrails for what it must not do, and you attach playbooks and skills. A step where a person decides holds no agent: a reviewer looks at the output of the preceding step and decides. Each run leaves a run record of what it did.
| Term | Meaning |
|---|---|
| Project | The unit that groups steps and agents. Can be classified with tags |
| Step | Where you place an agent. One step can hold several. A review step holds none |
| Agent | The worker. A goal (instructions) and guardrails define how it behaves, and playbooks and skills are attached to it. Reusable across steps and projects |
| Playbook | The procedure you can attach to an agent, at most one. It sets out the steps of work you repeat, each with its purpose, guidance and success criteria. The agent adjusts the steps to suit the request |
| Skill | A way of carrying out one particular piece of work. An agent can have several. You register a document describing the method, together with scripts or other files if it needs them |
| Placement | Each single agent placed on a step. Place the same agent on two steps and you have two placements. In a cycle the schedule is set at this level too |
| Review step | A step where a person decides. A reviewer looks at the output of the preceding step and chooses one of three: advance to the next step, send it back, or reject it. The request that reaches the reviewer is a review request |
| Run record | The detail of a single run. It traces the request, the evidence, what was done, the confirmation and the result. The list of runs is the “Execution history”; the record of operations inside a project is “Records (audit trail)” |
| Execution rules | The setting that decides which operations an agent may perform. A project inherits the execution rules set for the whole tenant and can only tighten them |
| Ledger | A connector the project's agents read from and write to as a shared record. You designate one by marking a connector “Use as ledger”. Marking it does nothing on its own: what is read and written there, and how, is settled by each agent's goal (instructions) |
One more word comes up when you create an agent: Claw. A Claw is tied to a conversation and runs on a schedule you set. It is not something you use in a project — it appears only as something to copy an agent's contents from. Only the contents are copied; triggers such as the time are not carried over.
The two types
You choose the type when you create a project. It cannot be changed afterwards.
| Cycle | Workflow | |
|---|---|---|
| How it starts | Each placed agent runs on a schedule you set for it. A webhook or email can also start it | You start it manually. It can also start at a set time, from the external API or by email. When it starts at a set time, you can choose which step it starts from |
| How it progresses | Agents do not wait for each other. Each runs on its own schedule | One run advances through the steps. A step begins after the previous one finishes. It can also divide into lanes partway and go forward in them at the same time |
| How it ends | It does not end. It keeps running until you stop it | Each run ends — Done, Rejected, Failed or Cancelled |
| Examples of work it suits | Picking up incoming requests and giving a first response. Checking the state at a set interval and raising an alert when something is wrong | Taking one application, doing the work, and finishing it through a reviewer. Running once at month end to produce a report |
How you use it
Choose the work
Decide what to hand over, what counts as done, and where the results go.
Decide what to connect to
Add the external services this project uses, and authenticate them.
Decide how much to delegate
Lay out the steps, place the agents, and decide which operations are allowed.
Use it with review
In a workflow, a person decides at a review step and moves it on.
Revisit with the record
Look at the track record, and revisit what to fix and how reviews are set.
Where to read at each stage.
- Choose the work — how much to hand over, what counts as done, where to divide projects, and where the results go (Design Guide)
- Decide what to connect to — Connection Guide
- Decide how much to delegate — type, steps and agents (Design Guide), which operations are allowed (Execution Rules Guide), how it starts (Trigger Guide)
- Use it with review — where people decide (Review Guide), who looks where (Record Guide), how to stop it (Trigger Guide)
- Revisit with the record — read the track record, choose what to fix, try it before applying, and compare against when the work is done (Record Guide)
Where things are in the UI
A project has a row of tabs. “Workflow history” is for workflows only, so a cycle has three.
| Tab | What you see |
|---|---|
| Overview | Metrics for a period you choose (KPI) and a diagram of the agents placed on each step. A workflow also shows the latest status of each step; a cycle, what each agent has been doing (see the workflow runs in progress under Workflow history). The purpose, members, registered agents, connections and execution rules are gathered under “Project info” |
| Execution history | The list of runs. Opening one shows its run record. The record of operations inside the project (Records (audit trail)) is here too |
| Workflow history | Recent workflow runs, including finished ones |
| Settings | Basic information, Triggers and notifications (Triggers in a cycle), Connectors, Members and Execution rules |
Steps and agents are edited in a separate editor screen, not in the tabs. An owner sees “Edit” at the top of the project screen.
Projects can also be used through the external API (external API documentation).
How it works
What the screens do not tell you is collected here. Read this before you start configuring.
- Projects and steps — a step is where you place an agent. A review step holds no agent. A playbook is the procedure and a skill is the method; both belong to the agent
- Cycle and workflow — in a cycle each placement runs independently. The type cannot be changed afterwards
- Who decides what — admin roles and project roles are different things. Even an owner cannot decide a review unless they are named as a reviewer (only a review request raised from an agent's result has, as a rule, every owner as its reviewers)
- Human review — a review step decides whether the work moves on. It is not a mechanism for stopping an agent mid-run
- How much to delegate — execution rules can only tighten, never loosen. Where they apply, and where your own responsibility begins
- Records and metrics — read a run record through five questions. The agent completion rate is not a measure of business outcome
Quickstart
Workflow
Create the steps, set the reviewers, and run your first workflow.
→Cycle
Place agents on steps, set the schedule they run on, and get the first sweep going.
→Guides
- Design Guide — what to settle for the work, how to divide it, and how much to delegate to each step and agent
- Execution Rules Guide — which operations an agent may perform. How the global basic rules are inherited, and how to write forbidden operations
- Review Guide — who decides, and where. How reviewers are chosen, and what happens when the deadline passes
- Trigger Guide — when it runs and how to stop it. Choosing between a set time, a per-agent schedule, a call from outside, and manual
- Connection Guide — what to connect to, and where to notify
- Record Guide — how to read the results, and what to revisit from the track record
References
- Settings Reference — the fixed values: number of steps, number of skills, deadlines, how to write times
- Run Results Reference — what each displayed result means, and the formula for the agent completion rate
- Records Reference — the events that are recorded, the CSV columns, and what an export contains
- Glossary — how the words on screen map to the words in records and settings
- Roles Reference — what owners, members, organization admins and system administrators can each do