Building control systems for autonomous agents.
DAG orchestration · MCP-native tooling · runtime routing · evidence-gated recovery
I work on the runtime layer between a model and reliable execution: routing work into bounded DAG lanes, coordinating multiple coding agents, capturing evidence, and recovering without losing operator control.
|
Provider-neutral control plane for Codex, Claude Code, OpenCode, and local coding agents. Routes tasks into scoped DAG lanes with replayable evidence.
|
Public MCP wrapper for topology-aware multi-agent orchestration, synthesis, and consistency verification.
|
|
Updated MCP server for agent-driven Roblox Studio workflows.
|
DAG-based orchestration engine and product layer for multi-agent LLM workflows.
|
| Work | Publication | Status |
|---|---|---|
| AdaptOrch: DAG-based Multi-Agent LLM Orchestration | arXiv:2602.16873 | IEEE ISAIA 2026, under review |
| Cognitive State Engineering | TechRxiv | Preprint |
| CoCo Theory | Zenodo | Preprint |
Current focus: multi-agent systems, LLM orchestration, evaluation, feedback loops, and control-loop-inspired AI design.
- DAG topology routing, cycle detection, and bounded parallel execution
- Reasoning-content handling across Kimi, Z.ai, and DeepSeek
- Tool-use evaluation loops, replayable evidence, and regression testing
- Telegram gateways, hooks, and proactive agent scheduling
Currently building the runtime layer for autonomous multi-agent systems.