Building the coordination, context, and decision layers that let teams of AI agents work like teams of people.
Twelve years of coordinating large human teams taught me that the hard part is never one smart worker — it's keeping many workers in sync, informed, and making good collective decisions. AI agents have exactly the same problem. That's what I build at night.
Three projects, one thesis: multi-agent systems fail at the seams — agents step on each other, work from stale knowledge, and make decisions nobody can audit. Each project owns one seam.
flowchart LR
subgraph Agents["Parallel AI agents"]
A1[Claude Code] --- A2[Cursor] --- A3[Copilot / Devin]
end
B["📚 Braxis<br/><b>Context layer</b><br/>what agents know"] --> Agents
Agents <--> N["🔀 Neo<br/><b>Coordination layer</b><br/>how agents share work"]
Agents --> AU["⚖️ AURA<br/><b>Decision layer</b><br/>how agents decide"]
| Layer | Project | Question it answers |
|---|---|---|
| Coordination | Neo | Two agents want to edit the same code — who goes, and does the other one know what changed? |
| Context | Braxis | Is the project knowledge my agents read still true today? |
| Decision | AURA 🔒 | When agents disagree on a high-stakes call, how do we decide — and prove why? |
When several Claude Code sessions work in the same repo, they can't see each other. Neo gives them a shared nervous system — it runs as Claude Code hooks, so every prompt, read, edit and shell command passes through it.
- Pre-generation intent checking — your prompt becomes declared intent, and each edit is rated LOW / MEDIUM / HIGH risk against everyone else's work before code is written
- Region-scoped locks + FIFO queue — the first editor takes the lock; others queue and are auto-promoted when the holder finishes
- Staleness detection + delta-only refresh — the differentiator. If a file changed after an agent last read it, Neo blocks the edit and refreshes only what changed instead of a full re-read. A broad prior-art sweep (GitHub, arXiv, PyPI, npm, HN) found no other coordination tool combining the two
- Breaking-change & blast-radius gates — flags changed signatures, lists downstream importers, and blocks edits while a caller is out of sync
- One-script install into any repo:
bash scripts/install_neo.sh <your-repo>
🤖 Neo was built almost entirely by Claude Code — a coding agent building the coordination layer it needs to work alongside other copies of itself.
Roadmap: intent-to-execution verification, multi-generational context lineage, consensus-based conflict resolution, deadlock detection, and a free GitHub App.
Agents read AGENTS.md, CLAUDE.md and .cursorrules to understand a codebase — and those files go stale the moment the code moves. Braxis generates them from static analysis and keeps them in sync on every push.
pip install braxis
braxis score # how agent-ready is this repo? (0–100)
braxis generate # AGENTS.md · CLAUDE.md · .cursorrules · .agentic-config.json- Agent-readiness score across architecture, testing, conventions, security, docs and more — with score history to track improvement
- Framework detection in three tiers (fast pattern / AST / deep wrapper analysis), 10+ languages
- CI-native — a GitHub Actions workflow regenerates context and opens a PR when your code drifts
- Zero dependencies, single-file core; optional Claude-powered recommendations
- Eats its own dog food: Braxis scores itself and the test suite grew out of its own recommendations
Where it fits: Neo keeps agents coordinated — Braxis keeps what they know current.
Most agent frameworks let one model decide, or let agents chat until they agree. AURA's Council Protocol (ACP) treats a decision like a well-run governance body: specialized roles — Memory, Guardian, Learning, Specialist, Shadow, Experimental — each take a stance, vote in two phases, deliberate under a budget, and leave an immutable record of every dissent.
- Validated on real public data — on IBM Telco churn (7,043 customers, independent ground truth), Guardian dissent carried a 19× lift; on Credit Card Fraud (284,807 transactions), stance-based Guardian lift reached 15.74×
- Throughput: 206K council decisions in ~9 seconds on Instacart-scale data
- Domain-agnostic kernel with adapters for customer lifecycle, fraud, healthcare, DevOps/SRE and more — 9 templated domains
- Cross-model tested with real LLM calls (Claude Sonnet 4.5, GPT-4o, Qwen, Llama) plus a blind grader
- Research-grade honesty — the paper documents its own boundary conditions and negative results (where deliberation doesn't help), not just the wins
📄 54-page paper published as a Zenodo preprint (DOI), now in journal review. Companion work in progress: a four-LLM comparison study and spin-off papers on self-auditing ML evaluation, kernel architecture, healthcare and fraud triage.
Vision — AURA OS: ACP · AURA Lifecycle · AURA Memory · AURA Search · AURA SDK.
| Project | What it is |
|---|---|
| AisleGlyph | CPU-first framework for retail-specific tokenization — mines domain phrases, builds a reversible prompt codec and a small BPE tokenizer, and reports token savings |
| PAAP · v2 | Privacy-Assured Anonymous Participation — blind-issued credentials and scoped nullifiers for eligible-only, one-time actions without identity login |
| BrowseAI | Chrome extension for memory-first browsing — session and context history that persists. Live on the Chrome Web Store |
| GroupBuy | Group-discount buying platform |
| Snake: Genesis Shift | A remastered take on the classic |
| Project | What it is |
|---|---|
| AppResilience | Automated API resilience testing for mobile apps — auto-capture traffic, generate response variants (missing fields, type mismatches, delays, errors), and report pass rates |
| RetailQuest | Gamified convenience-retail loyalty concept |
| Penguins Plunge | Cross-platform game built with Kotlin Multiplatform |
| PageHugger · TheBeerClub · Rules Forge · Nova Paint 7FAD | Private side projects in active development |
Everything I build ships under Primeape94 Lab — my personal signature across apps, sites, repos and AI tools.
Leading teams by day. Coordinating agents by night.