Spec

Chain
conversations

Sequential multi-bot collaboration with
built-in turn control.

Problem

Why not group chats

Group chats cause simultaneous replies from all participants. Turn-taking requires constant moderation. Bots can't resist replying even when told to hold.

Concept

What is a chain

A chain is a sequential conversation thread where:

  • One bot is active at a time
  • Each new participant reads the full chain history before contributing
  • CEO (or a coordinator) controls who speaks next by chaining them in
  • No simultaneous replies because only the active bot receives the "your turn" signal
Flow

How it works

1
Start the chain

CEO creates chain monitor-ia-redesign and posts the brief from Max.

2
Chain in researcher

Researcher receives full chain history + prompt "research dashboard IA patterns." Does the work, posts findings. Turn ends, chain returns to CEO.

3
Chain in designer

Designer receives full chain history (Max's brief + researcher's findings) + prompt "build a proposal." Produces the proposal, posts to chain. Turn returns to CEO.

4
Close the chain

CEO posts summary and outcome. Chain is archived, viewable in monitor UI.

Properties

Key properties

Sequential

Only one bot active at a time. No race conditions.

Full context

Each bot reads the entire chain before contributing.

Coordinator-controlled

CEO decides who speaks next and what they should do.

Auditable

The chain is a complete record of the collaboration.

Comparison

Chain vs Group Chat vs DM

Feature
DM
Group Chat
Chain
Participants
2
N (simultaneous)
N (sequential)
Turn control
Natural
Needs moderation
Built-in
Context
None shared
Everyone sees all
Full history on entry
Race conditions
No
Yes
No
Best for
Quick 1:1
Real-time with Max
Multi-bot collab
Implementation

Options

Option A
Convention-based

CEO manually DMs each bot in sequence, copying relevant context. No code needed. Chain history lives in CEO's memory.

Works today. Error-prone, not viewable in UI.
Recommended
Option C — Hybrid

Trial with manual chaining (Option A) to validate the pattern. Build colony tools (colony_chain_create, colony_chain_add, colony_chain_message) once we're confident chains work better than groups.

Fast to start, informed by real usage.
Option B
Colony feature

New colony tools, chain history in DB, bots receive history automatically, viewable in monitor dashboard as a thread.

Clean and auditable. Needs implementation work.
Trial

Next step

Pick a real task and run it as a chain:

  1. CEO starts with a brief
  2. Chain researcher for investigation
  3. Chain designer for proposal
  4. Relay result to Max

Track: did it avoid simultaneous replies? Was context sharing smooth? What was awkward?