Autonomous agents · Building now

AJO.

A message bus that routes work across a fleet of Claude agents and runs them as one system. 18 projects. 85 roles. One router.

The message bus — live trace
routing
AJOcanonical bus

Every message routes through the bus — never agent-to-agent — between Daniel and the eight subprojects below. Order is emergent, not scripted.

Rotate your phone to watch the router run live
task
reply
concern
sign-off
A live trace of message kinds routing through the substrate. Nothing runs in a fixed order — Tool Engineering can ask Agent Dev to build a new role mid-task, and any role can raise a concern straight to Daniel.

Give AJO an objective. It splits the work across roles, hands off between them, and a quality judge scores whether the job was actually done — not whether the run finished. The objective can change. The coordination underneath it doesn't.

01How it works

Built in layers, inert by default.

  • One router, many roles — a file-canonical, SQLite-backed message bus routes work across 18 projects and 85 roles, each a scoped agent with its own charter.
  • A campaign conductor — decomposes an objective into a dependency graph and checks each step’s preconditions before unblocking the next — a failed check inserts a repair instead of plowing ahead.
  • Per-role memory at spawn — each agent gets exactly the context its role needs, pushed in when it wakes.
  • A quality judge — scores whether the work actually satisfies the task, so "the run completed" is never mistaken for "the job is done."
  • Quarantined input + safe-ops — untrusted outside input stays walled off until a human approves the pilot. A kill switch hard-stops every write. Each layer sits inert by default, gated behind that approval.
02What it runs on

A self-driving discovery loop.

Discovery runs on two engines working as one. The Researcher decides what's worth asking. The Pain Point Pipeline answers from real captured pages, weighted by proof that someone pays. The Researcher reads the evidence and asks sharper. Point it at a domain and it drives itself toward a verdict on a buildable pain.

03What was hard

Making a fleet trustworthy.

The hard part was never getting one agent to act. It's getting many of them to act reliably. Runaway-loop failure modes are killed structurally — a derivation-depth cap and atomic claim primitives, not a prompt asking nicely. The quality judge needed its own recalibration, to tell a bad scorer apart from bad work. And nothing runs live until a human arms it.

Code is private. This page is the public record of the system.

04The roles

One router, eight fronts.

Every one of these runs as a scoped role inside AJO, coordinated by the same message bus — working through whatever comes up, not a fixed script.

Orchestrator
Assigns work across roles and projects — holds the message bus together.
Discovery
Scopes a problem before any agent touches it — the front door into AJO.
Pain Point Pipeline
Discovery’s live engine — surfaces and qualifies the pain points worth pursuing.
Researcher
Decides what’s worth asking, ranked by value of information.
Tracking
Sweeps for open priorities and keeps the long-horizon queue visible.
Agent Dev
Scaffolds a new role — prompts, guardrails, eval harness.
Tool Engineering
Builds and hardens the typed tools agents call.
Systems
State, storage, and the plumbing everything else sits on.
← All work