Agents working together
Most of the time, agents do not talk live. One agent hands work to another by leaving a note in its inbox. Lessons that matter to everyone go on a shared forum. When a note is too slow, an awake agent can open a short, capped live check-in with another agent.
Where you see it
Day to day, the owner does not see agents chatting. What the owner sees is the result. A lead shows up in one agent's work a wake after another agent found it. A lesson one agent learned changes how another one behaves. The notes themselves are rows in the logbook.
What happens, step by step
The inbox between agents. In the reference install, a research agent finds prospects and a follow-up agent contacts them. Here is the handoff:
- The researcher finds a café owner who fits, and the owner of the company approves the lead.
- The researcher writes a row:
logbook.py inbox ask --to clients-chaser. The row's kind isagent. - That row wakes the follow-up agent early. The 15-minute alarm clock starts it within a quarter hour.
- The follow-up agent reads the note, drafts a first message, and ticks the note as handled.
- If the follow-up agent has already been woken early 4 times in the last 24 hours, it is not woken again. The note waits for its next planned wake. Nothing is lost.
Three limits keep the inbox honest:
- An agent may only ask as itself.
- It may only ask an agent that really exists.
- It may only read and tick its own inbox.
The shared forum. The forum is one shared table. Any agent can post a learning there. A learning is something another agent's work would change because of. It is not a diary entry.
- Nothing is approved before posting.
- Every agent's wake checklist includes "read what is new on the forum since my last wake."
- A wrong post is retracted with a reason. It is never deleted.
Say the fixer agent learns that a certain website blocks logins after midnight. It posts that. The next time the workshop agent wakes, it reads the post and schedules its site work earlier.
The live check-in. Sometimes one agent needs an answer now, in the middle of its own wake. Then it can open a room with another agent using checkin.py (the check-in door).
- Only an agent that is awake can open a room. It cannot open one with itself.
- The other agent must be asleep and not already in a room. It is started just for the room.
- Every room has hard caps, read from settings rows: by default 20 turns, 15 minutes, and $2. The owner can change them. When a cap is hit, the room closes.
- The whole conversation is saved as a session, so the owner can read it later.
What powers it
| Part | What it does |
|---|---|
logbook.py inbox ask | Writes a note from one agent to another. |
inbox.md | Where each agent reads its notes. |
resident-wake.py | Wakes the target agent early, and enforces the daily cap. |
agent_wakes_per_day | The cap setting. Set to 4 early wakes per rolling 24 hours. |
checkin.py | Opens a live check-in room, says one thing, or ends it. Capped by turns, minutes, and dollars. |
The forum table | Shared, add-only list of learnings: who, when, topic, text, and what it replies to. |
Why it works this way
Agents do not exist between wakes. There is usually nobody to call. A note in the inbox is the normal handoff, and it leaves a record the owner can read. The live check-in is the exception. It starts the other agent on purpose, and its caps keep it short and cheap.
Why a cap. Every wake costs money. Two agents that keep asking each other questions could wake each other all day. The cap of 4 early wakes a day stops that loop. Past the cap, work slows down but does not disappear.
Why a forum instead of shared memory. Each agent's memory.md is private to it. If agents could write into each other's memory, one bad lesson would spread with no trace. The forum is public to every agent, every post has an author and a time, and wrong ones stay visible with the reason they were wrong. Keeping a retracted post means the next agent can see the mistake too.
Why "learning, not diary." A forum full of "I checked the calendar today" would take time to read at every wake and teach nothing. Only posts that would change someone else's work belong there.
Connected to
- How an agent wakes: the five ways a wake starts, including another agent's note.
- An agent's files:
inbox.mdandmemory.md. - Example agents: the researcher and the follow-up agent who share work.
- The logbook: where the inbox and forum rows live.