Skip to content

Your first project

This walkthrough takes a small project from nothing to one merged change. It uses three agents: one plans, one builds and one reviews. The screenshots come from doing exactly this on a small invoicing service. It takes about fifteen minutes, most of it the agents working.

You need AgentWeave installed and at least one agent CLI signed in.

1. Open the project

cd ~/code/northwind-invoicing
agentweave

The app opens on the project's Overview. Your projects are listed in the left rail, and each one expands to show its agents and their conversations. Across the top are the project's tabs: Overview, Tasks, Spec, Vault, Jobs, Activity. The gear on the right opens the project settings.

Set the main branch

Open Project settings → General → Integration and check that Main branch names the branch approved work should merge into. AgentWeave adopts the repository's current branch when it opens a project. You can also add Checks there, such as pytest -q. Approval is then refused until the checks pass on the merged result.

2. Add three agents

Click + next to the project in the rail (Add agent). An agent needs a name, a provider, a model, and optionally a charter:

Name Provider and model Charter
planner Claude Code · Sonnet Spec Author
builder Claude Code · Sonnet Developer
reviewer Claude Code · Sonnet Code Reviewer

The Create agent dialog with name, provider, model and charter The Create agent dialog with name, provider, model and charter

A provider whose CLI cannot launch is greyed out, with the reason. The Hub creates the matching runner for you. Everything else, such as default permissions, timeouts and grants, lives in the agent's settings, and the defaults are sensible.

3. Ask for a change

Open planner and describe what you want. Be concrete about the outcome, not the implementation:

Two things for Northwind Invoicing. 1) The CSV export should end with a totals row: subtotal, tax and total for the invoice region. 2) Tax must round half-up to the cent, per region. Please author a short change specification with requirements and scenarios, then the tasks for the builder, ordered by dependency.

The planner reads the code and writes a change specification into spec/changes/…/spec.json. The Hub gives each requirement a stable identifier (FR-1 … FR-4). Open it beside the conversation from the side panel's Specs tab, or on the Spec tab:

A conversation with the planner, with the change specification it wrote open in the side panel A conversation with the planner, with the change specification it wrote open in the side panel

If the agent needs a decision from you, it asks with a question card above the composer, and its turn waits for your answer. Agents never guess at what you meant.

4. Check, approve and build

On the Spec tab, Check lists anything still missing, exactly as the agent sees it, and moves nothing: the agent keeps writing while you read. Approve approves the version you read (if the agent changed it since, the Hub says so and you reload) and lays the tasks out as a plan; nothing starts. Build… then starts building it the way the document recommends, now or later, and Approve and build… does both at once. If the Hub found gaps, for example a scenario with nobody named to check it, approving lists them first. You can fix them, or press Approve anyway. The gaps you overrode are recorded on the document.

The approval panel listing gaps, with an Approve anyway button The approval panel listing gaps, with an Approve anyway button

Agents can draft and edit a specification, but only you can approve it. An agent asked to build from an unapproved specification says so and stops.

5. Build

Open Tasks. Approval put the document's tasks on the board, in dependency order. Hover over a card and choose Start builder on this, or tell builder in its conversation to take the task.

The builder works in the task's own worktree (.agentweave/tasks/<task-id>, on branch agentweave/task/<task-id>). It implements the change, runs the tests, commits, and records evidence against the requirement: the command, its output, and the commit. Then it moves the task to completed. That is the last thing you had to start: the review starts by itself.

The task board with tasks in Pending, Completed and Approved The task board with tasks in Pending, Completed and Approved

A task that depends on another cannot start until its prerequisite is approved. Completed is not enough.

6. Review and merge

You do not need to ask for a review. Once the task is completed and the project's checks pass, the Hub starts a review turn by an agent other than the builder, and the task moves to under review. It picks the task's named checker if it has one; otherwise it chooses an idle agent other than the builder, which may be planner rather than reviewer, and the task's story says who was picked and that nobody was named. If the builder is the project's only agent, no review starts and the task says there is nobody else to review it.

To choose the reviewer yourself, hover over a completed card and choose Review with reviewer. The menu lists every agent except the task's author.

A review turn gets a clean checkout of exactly the commit the evidence names. The reviewer runs the tests, decides the evidence, and either approves the task or sends it back to the builder with the reason. The builder cannot approve its own work.

Approving merges. The task's evidenced commit is merged into your main branch, locally, in the same step, and nothing is pushed. If the merge would conflict, approval is refused and names the files. Open the task to see its whole story:

A task's story: state, cost, runs, evidence and the merge, in time order A task's story: state, cost, runs, evidence and the merge, in time order

$ git log --oneline main
ab4099c Integrate INV-1: Round tax half-up to the cent
747084d Round tax half-up to the cent (FR-2)
888d46d Initial invoicing service

Land it is your override

A completed task's drawer also offers Land it, which approves and merges the task without a review. It asks you to confirm first, and says how many pieces of evidence nobody has reviewed. Use it when you mean to skip the review, not as the normal way to merge.

Where to go next

  • The app: every screen, and what it is for.
  • Specifications: the authoring journey, rigor, and folding a shipped change into current behaviour.
  • Jobs, loops and flows: let a flow work through an approved specification on its own, with every task reviewed.
  • Permission postures: what an agent may do before it has to ask you.