| name | nate-b-jones-orchestration-intelligence |
| description | Designs, audits, and optimizes multi-agent orchestration using Nate B. Jones's DPVI pattern (Decompose → Parallelize → Verify → Iterate) — the convergence architecture independently arrived at by Anthropic, Google DeepMind, OpenAI, and Cursor. Treats harness design (memory, task files, progress tracking, restart procedures) as the primary determinant of agent success, not model intelligence. Classifies work by domain verifiability to determine safe delegation. Use when designing a new multi-agent system, coordinating >3 agents, agents are capable but inconsistent (suspect harness issues), deciding what to delegate vs. retain for humans, or agent systems show diminishing returns. Trigger proactively whenever the user says "coordinate agents", "multi-agent", "swarm", "parallel agents", or "why are my agents inconsistent". For memory/context architecture use nate-b-jones-context-engineering; for self-improvement loops use nate-b-jones-auto-improvement-loops. |
| expert | Nate B. Jones |
| domain | AI/Automation — Multi-Agent Orchestration |
| version | 1 |
| source | "4 AI Labs Built the Same System Without Talking to Each Other" (YouTube, 2026) |
Nate B. Jones — Orchestration Intelligence
The jagged frontier of AI is not a property of intelligence — it's an artifact of missing organizational structure. Apply human team management principles to agents, and the frontier smooths.
Core Capability
Design, audit, and optimize multi-agent orchestration systems using the DPVI pattern — the convergence architecture independently built by Anthropic, Google DeepMind, OpenAI, and Cursor. This skill treats harness design (memory, task files, progress tracking, restart procedures) as the primary determinant of agent success, not model intelligence. It classifies work by domain verifiability to determine safe delegation levels.
What This Skill Produces
| Workflow | Output |
|---|
| Orchestration Architecture Blueprint | Complete multi-agent design with planner-worker-judge hierarchy and verification loops |
| Harness Design Audit | Agent scaffolding evaluation with gap analysis and fixes |
| Domain Verifiability Mapper | Work classification into machine-checkable / expert-checkable / unverifiable tiers |
| Sniff-Check Protocol Builder | Evaluation criteria and quality gates for any domain |
| Bloat-to-Depth Optimizer | Multi-agent simplification plan that strips overhead while preserving capability |
| Orchestration Telemetry Feedback | Self-optimizing layer — instruments coordination decisions, detects inefficiency patterns, feeds retrospectives into future runs |
| Worker Isolation + Merge Infrastructure Design | Two-tier scale architecture: isolated worker contracts, merge/refinery layer, episodic state — avoids coordination collapse past small agent counts |
When to Deploy
- Designing a new multi-agent system or coordinating >3 agents
- Agents are capable but inconsistent — suspect harness issues
- Deciding what work to delegate to agents vs. retain for humans
- Agent system growing complex and showing diminishing returns
- Building evaluation criteria for AI-generated work in any domain
- Multi-agent pipeline runs repeatedly and should get more efficient over time
Stacking Opportunities
- + Nick Saraev (Agentic Workflows): Self-annealing + harness design = self-correcting agent systems
- + Oren (Taste Mastery): Sniff-check protocols + CEV evaluation = formal quality layer for agent outputs
