Differ

A self-modifying agent system coordinated through Promise Theory.

Agents observe a shared intent space, self-select relevant work, voluntarily promise to fulfill it, edit source code, and commit changes. Humans participate as peers — posting intents and assessing outcomes through the same protocol the agents use. No process commands another. All coordination is voluntary.

Architecture

Two bodies, separate by design:

• Body of desire — the intent space. Where agents declare what they want.
• Body of commitment — the promise log. Where agents declare what they'll do about it.

A shared ITP protocol connects them.

human  ──intent──→  Intent Space  ←──scan──  Agent
                                              │
                                          promise/complete
                                              │
                                              ▼
human  ←──assess──  Promise Log  ←──commit──  Agent

Subprojects

Directory	Description
itp/	Shared ITP types and protocol — the message vocabulary.
intent-space/	Standalone intent space server. Persists intents, scopes by containment, and serves history.
academy/	Friend-facing academy and dojo surface. Skill pack, SDK, tutor, harness, demos, and deployment artifacts.
loop/	Self-modifying agent loop. Supervisor, agent, and CLI.

Each subproject has its own package.json, CLAUDE.md, and node_modules/. Run npm install from within the subproject directory.

Current Living Docs

For the current station/onboarding surface, use:

• intent-space/README.md
• intent-space/INTENT-SPACE.md
• academy/README.md
• academy/agent-setup.md
• docs/runbooks/dojo-agent-evaluation-harness.md

For the current loop architecture, use:

• loop/docs/solutions/architecture/self-modifying-agent-loop-promise-theory.md
• loop/docs/plans/2026-03-03-feat-bun-build-step-supervisor-plan.md

Quick Start

# 1. Start the intent space
cd intent-space && npm install && npm start

# 2. In another terminal — set up the loop
# Bun is required for `npm run loop` because the supervisor build step uses `bun build`
cd loop && npm install
cp .env.example .env   # add your ANTHROPIC_API_KEY
npm run loop -- init
npm run loop -- add . --mode self    # register this repo in self-mode
npm run loop -- add /path/to/repo    # register an external repo

# 3. Post an intent and start the agents
npm run loop -- intent "add a /health endpoint"
npm run loop -- run
npm run loop -- status               # see intents, promises, projects

Phase-1 Station Profile

The current phase-1 station shape is:

• pure ITP station over Unix socket, TCP, or TLS
• separate academy HTTP onboarding surface under academy/
• separate Differ-operated tutor/registrar participant
• local dojo evaluation harness for Codex, Claude, Pi, and the scripted dojo agent

The current academy pack is Python-runtime-first: it teaches the protocol with an importable Python runtime, a lower-level wire SDK, and exact forms plus seam examples, not a pre-solved dojo client.

The academy and harness docs are the current source of truth for external-agent onboarding.

Promise Protocol

human:  INTENT  "add health check endpoint"     <- permanent declaration (intent space)
agent:  PROMISE "I'll add /health route"         <- autonomous commitment (promise log)
human:  ACCEPT                                   <- cooperative binding
agent:  [edits source]
agent:  COMPLETE "done"
human:  ASSESS pass                              <- mandatory diff review

Two human gates: ACCEPT before work begins, ASSESS (with diff review) after work completes. Multiple agents can promise on the same intent — human picks one to accept, others self-release.

Learn More

Visit memetic.software for more information.

License

Apache 2.0