3 · Two agents at once

Fifteen minutes. One agent builds; another reads its work and says what is wrong. You will also learn to read two conversations on one screen.

Before you start

Put a loop under it. Two CLIs logged in makes this more interesting, but one is enough — two agents of the same CLI are still two agents.

Step 1 — fetch the official flowverse

The interesting flows are not in the package. They are in a flowverse — a git repository of flows — and official is listed from the start whether or not it has been fetched.

In hmz:

/flow

Press → until the tab says official. If the list is empty or stale, press ctrl+r to fetch it. It keeps drawing while it clones, and says underneath what became of it.

Adding a flowverse is trusting that repository with this machine

A flow is a Python file, and listing what a flowverse holds imports every file in it. official is humanfia/flowverse. See Security.

Step 2 — take rlar

official/rlar

rlar drives two agents with names, and the names matter:

Agent	What it is for
actor	works in one session and must remember
reviewer	arrives fresh each round and must not

The review is the actor's next prompt, word for word. And the reviewer is also the one that says the task is finished — which is what ends the run.

Step 3 — set up each agent

Choosing the flow walks you straight into /agents, one agent at a time, by the name the flow calls it. Each agent is three steps:

1. Which CLI, and which account. A tab per CLI that is installed; as installed is the first row. (Accounts are providers — tutorial 16.)
2. Which model, and at what effort. Type a few letters to narrow. ←/→ adjusts the effort.
3. Where it works — only for an agent the flow says may be pointed at a machine. rlar does not, so this step does not exist here.

esc is always the step before, and off the first step of the first agent it leaves the walk entirely, changing nothing.

Give the reviewer a different CLI if you have one. Give it the same one if you do not — two agents at one configuration are still two agents, and that is the point.

Make the reviewer read-only

On the reviewer's model step, ctrl+p → read-only. Now it can read the change and cannot touch it. See Permissions.

Step 4 — run it

Work through TASK.md.

The line above the editor now has two rows:

   actor · claude/claude-opus-5:max · ● 1 of 1
   reviewer · codex/gpt-5.6-sol:high · ○ 3 · unread

●	that agent has a turn open
○	it has stopped
1 of 1	which conversation of that agent's you are reading
3	the reviewer has opened three, one per round
unread	it has said something since you last looked at it

Step 5 — read the other conversation

tab.

The transcript writes a line saying which conversation is being read from there down, and draws what it has said under it. shift+tab goes back; both wrap.

Three things to know:

• They step between the ones that are working. A conversation between its turns is read once you are on it, but is not stepped onto.
• Nothing is taken off the screen. A conversation ending under you says that conversation has gone; only /clear clears.
• A line you type reaches the conversation you are reading — not whichever agent happens to be working. Say something to the actor while reading the reviewer and you have said it to the reviewer.

See Many conversations at once.

Step 6 — look at the handover graph

/status

Now it says something: every handover between agents, with how often it happened. A two-agent loop that was supposed to alternate and is in fact one agent doing everything looks different here from the first glance.

Step 7 — the same thing without the interface

hmz exec -f official/rlar \
    -a claude/claude-opus-5:max \
    -a cli=codex,model=gpt-5.6-sol,effort=high,permission=read-only \
    "$(cat TASK.md)"

One -a per agent the flow drives, in the order it takes them. Get the count wrong and it is refused before any agent runs:

$ hmz exec -f official/rlar -a claude/claude-opus-5:max "fix the build"
hmz exec: error: official/rlar: the flow drives 2 agents, 1 given

That check exists so a two-agent flow started with one fails now rather than on an unpacking hours into a loop, with a turn's work already behind it.

Other two-agent flows to try

official/flame_chase	two agents take turns on the same task; each reads the repository, not a history
official/humanize1:rlcr	builds a plan under review until nothing is left to say — run it in a git repository

What you now know

• A flow names its agents, and everything talks about them by those names.
• /agents is three steps per agent, and esc is always the step before.
• tab reads another conversation; a line goes to the one you are reading.
• -a is positional: one per agent, in the flow's own order.

Next

Run it unattended.