An agentic workflow, built and run: coder, reviewer, send it back

What an agentic workflow looks like

A chat with one AI is a loop you drive: you ask, read, correct, ask again. An agentic workflow moves the checking into the system. The steps are decided up front; the agents run them.

1. Work: a coding agent makes the change on its own branch.
2. Review: a second agent reads the change against the original brief and gives a verdict, OK or BLOCK, with findings.
3. Decide: OK goes on to you (or to a merge or deploy step). BLOCK sends the findings back to the same agent, in the same session, to fix its own work.
4. Stop: after a set number of rounds the work ends as blocked instead of looping forever.

Build it

You need Swarm on a Linux machine with a model connected and a project for your code; the first guide covers that in a few minutes. Then say what you want:

Your AI saves this workflow. It is plain JSON, kept with its history, and you can read or change it any time:

{
  "name": "review-coder-work",
  "project": "tally",
  "scope": { "role": "coder", "folder": "/srv/tally" },
  "when": "succeeded",
  "steps": [
    { "id": "review", "type": "agent", "worker": "reviewer",
      "brief": "Check the change against its brief: does it do exactly what was asked, does test.sh cover it, and is anything risky or out of scope? Give an OK or BLOCK verdict with concrete findings." },
    { "id": "send-back", "type": "reopen", "from": "review",
      "if": [{ "step": "review", "verdict": "block" }],
      "max_rounds": 2,
      "message": "The reviewer found problems. Fix them on the same branch:" }
  ]
}

Run it

Now any coder task in that folder goes through the workflow. We gave the coder a small fix in a bash tool called tally:

The coder worked on its own branch, ran the tests and committed: 35 model turns, $0.14 on Gemini 3.8 Flash. Its change:

-# FIXME: the total line is printed even for a single file
-printf '%-30s %8d %8d %8d\n' "total" "$total_l" "$total_w" "$total_c"
+if [ $# -gt 1 ]; then
+  printf '%-30s %8d %8d %8d\n' "total" "$total_l" "$total_w" "$total_c"
+fi

The workflow started the reviewer (Gemini 4 Argon) the moment the task ended. Its verdict came back 35 seconds later, quoted as written:

VERDICT: OK
SUMMARY: The commit accurately resolves the FIXME in tally.sh by guarding the "total" output line
with if [ $# -gt 1 ]; then ... fi. test.sh covers both the single-file case (asserting no total line
is printed) and the multi-file case (asserting the total line is printed). [...] Non-blocking note:
in test.sh:9, anchoring the single-file pattern as ^total would avoid a theoretical false positive.
FINDINGS: None

Because the verdict was OK, the send-back step was skipped, and the run record says why: condition not met: step review verdict is ok, not block. A BLOCK would have sent the findings into the coder's own session and started a second round.

Every run is on the record

swarm_list_workflows shows each workflow with its latest runs: which job it followed, each step's state and outcome, the reviewer's report and verdict, and when it started and finished. Ask "what did the reviewer say about the last change?" and your AI reads it from there.

Other steps you can add

• deliver: merge the work into the folder's branch, open a pull request, or save the result to a directory.
• run: your own script on the machine, for example to deploy or to post an alert. Only the machine's owner can add one.
• agent: any agent on the result: a finder to write a summary, a designer to check a page, or one of your own workers.
• if: run a step only when an earlier one succeeded, failed, or got a given verdict.