Your first flow: a research brief

#Goal

A flow where one agent researches a topic, another writes a short brief, and you approve it before it is used. Then you will make a rejection send the work back for another round.

#Prerequisites

• Two agents with working runtimes (or one: an agent can fill both roles). A model agent or a Read only CLI agent is ideal: this flow does not need to change files.

#1. Create the flow from the template

1. Flows → New flow (or Create your first flow).
2. Choose Research brief.
3. Name: QWERTY brief.
4. Who does what: pick an agent for Researcher and one for Writer.
5. Press Create flow.

Expected: the editor opens with four steps: Start → Research → Write the brief → Approve the brief, and the header says Ready to run.

#2. Look at the steps

• Click Research: its Task is filled in, and under Give it the output of the Start step is ticked, so it receives your run input.
• Click Write the brief: it receives Research's output.
• Click Approve the brief: Show the output of includes Write the brief, so you will see the brief when deciding.

Optional: click Start and add Instructions for every step in this flow, for example Write in plain English for a general audience.

#3. Run it

1. Press Run.
2. Input: the history of the QWERTY keyboard layout.
3. Press Run.

Expected: steps light up in order: Queued, Running, Completed. After Write the brief, the run pauses as Waiting for you, and a badge appears on Runs in the sidebar.

Behind the scenes: each agent step runs in the run's own private conversation. Hand-offs are off there, so the agents cannot wake each other; the arrows decide what happens next.

#4. Approve

1. Open Runs and select the run.
2. The approval card Approve the brief shows the output of Write the brief.
3. Optionally type Feedback, then press Approve.

Expected: the run ends Completed.

#5. Inspect it

Click Research on the run canvas:

• Input shows exactly what the agent received, including your input fenced as untrusted data.
• Output shows its reply.
• Metrics shows duration, queue wait, tokens and cost when reported.

Click the arrow from Research to Write the brief to see the value that moved along it. Try Replay to step through the run's events.

#6. Send rejected work back

1. Back in the editor, drag from Approve the brief's Rejected exit to Write the brief. The arrow turns dotted red: sends back.
2. Save, then Run again with a new topic.
3. When asked to approve, write feedback such as Shorter, and add a source for each claim, then press Reject.

Expected: Write the brief runs again as attempt 2, with your feedback in its prompt under What the reviewer asked you to change. You are asked to approve again.

Behind the scenes: a send-back reruns the target and everything after it, keeping earlier attempts for inspection. One arrow can send work back at most 10 times per run.

#Common mistakes

• Pressing Run with unsaved changes. Save first.
• Rejecting without a Rejected path. The run then ends Rejected. That is expected.
• Unticking inputs. An agent only sees outputs you tick.

#Next steps

• Run a flow from chat
• Steps reference