ollie/doc/MULTI_AGENT.md

Multi-Agent Coordination

ollie exposes agents as filesystem objects via the 9P protocol. This means multi-agent coordination requires no framework, SDK, or message bus — it is ordinary shell scripting against a mounted filesystem.

The Namespace

Every session lives under /s/{id}/:

Path	Purpose
/s/new	Create a session (write name, backend, model, agent, cwd)
/s/{id}/prompt	Submit a prompt (dispatched async on close)
/s/{id}/chat	Cumulative conversation history (append-only read)
/s/{id}/cfg	Session config (KV): state, backend, model, agent, cwd, params
/s/{id}/statewait	Blocking read — returns when state changes
/s/{id}/ctl	Control commands: stop, kill, compact, rn <name>

Utility files at the session root (/s/) provide higher-level operations:

File	Purpose
/s/b	One-shot: create, submit, wait, print result, kill
/s/bfg	Batch foreground: submit to existing session, wait, print
/s/bbg	Batch background: submit, return session path immediately
/s/sh	Interactive shell session

Global namespaces are shared across all sessions:

Path	Purpose
/t/{name}	Tool scripts (read/write)
/t/idx	Synthesized tool index (descriptions + args)
/a/{name}.json	Agent configs (supports mv, cp, rm)
/sk/{name}	Skill files
/p/{name}	Prompt templates

Spawning Agents

From within an agent session, use subagent_spawn:

subagent_spawn [-name NAME] [-agent AGENT] [-backend BACKEND] [-model MODEL] [-cwd DIR] <prompt>

This creates a child session, writes the prompt, and returns the session path (e.g. s/my-session). Child sessions carry the parent session ID in their name, so the spawning tree is recoverable from session IDs alone.

Coordination Patterns

Hub-and-Spoke (Conductor)

A root agent reasons about a task and delegates subtasks to workers. Workers report back via their chat history; the conductor reads /s/{id}/chat after statewait unblocks.

# Conductor spawns two workers
worker_a=$(subagent_spawn -model haiku "analyse /src/a.go")
worker_b=$(subagent_spawn -model haiku "analyse /src/b.go")

# Wait for both
until [ "$(cat $worker_a/state)" = "idle" ]; do cat $worker_a/statewait; done
until [ "$(cat $worker_b/state)" = "idle" ]; do cat $worker_b/statewait; done

# Collect results
cat $worker_a/chat $worker_b/chat

The conductor and workers can run on different models. Mechanical analysis goes to cheaper models; reasoning stays on the expensive one.

Pipeline

Agent A produces output; agent B consumes it. Composition via /tmp/ or direct path passing.

step1=$(subagent_spawn "extract all TODOs from /src/ and write to /tmp/todos.txt")
until [ "$(cat $step1/state)" = "idle" ]; do cat $step1/statewait; done
subagent_spawn "prioritise the TODOs in /tmp/todos.txt and produce a plan"

Scatter-Gather (Swarm)

Spawn N workers in parallel, collect when all reach idle. Because statewait blocks until state changes, no polling loop is needed.

# Scatter
for f in /src/*.go; do
  subagent_spawn "review $f for security issues" &
done
wait  # wait for all subagent_spawn calls to return session paths

Or from within execute_code using parallel steps:

{
  "steps": [
    { "parallel": [
      { "code": "subagent_spawn 'task A'" },
      { "code": "subagent_spawn 'task B'" },
      { "code": "subagent_spawn 'task C'" }
    ]}
  ]
}

Peer-to-Peer

Any agent can read any other agent's session directory. There is no privileged orchestrator role — agents coordinate by reading each other's chat and state files directly. A peer can wait on another peer's statewait, then act on what it reads in chat.

Reactive / Event-Driven

statewait is a proper synchronization primitive. A shell script (or another agent) blocks on it and wakes exactly when the target agent transitions state. This enables event-driven pipelines without polling:

# Block until agent finishes, then trigger downstream work
cat /s/worker-123/statewait
# agent is now idle — proceed

Self-Directed Spawning

Agents can spawn sub-agents mid-task without human direction. If an agent determines that parallelism or specialization would help, it calls subagent_spawn on its own. No human prompt is required to initiate a multi-agent workflow — the model decides.

Design Principle

ollie imposes no coordination model. There is no built-in conductor, swarm layer, or workflow engine. Orchestration is handled by the surrounding environment — shell scripts, cron, systemd, containers, or another agent.

This is intentional. The filesystem namespace is the API. Any tool that can read and write files can participate in a multi-agent workflow, which means the full Unix toolchain is available as an orchestration layer.