nate-b-jones-auto-improvement-loops

name	Nate B Jones — Auto-Improvement Loops
description	Design, deploy, and safeguard self-improving agent systems. The Karpathy loop applied to business harnesses — not just training code.

Nate B Jones — Auto-Improvement Loops

The magic isn't in the agent's intelligence — it's in the constraints. One editable file, one metric, one time budget. Build the foundations, or auto-improvement degrades your system faster than humans could.

Core Capability

Design, deploy, and safeguard self-improving agent systems using the Karpathy loop architecture. Distinguishes auto-research (optimizing model internals) from auto-agent (optimizing harness scaffolding — the universal pattern for every agent-deploying business). Covers the full lifecycle: triplet definition → readiness audit → meta/task architecture → trace infrastructure → emergent affordances → safety → deployment → system-level self-application.

What This Skill Produces

Workflow	Output
Karpathy Triplet Design	Three-part specification (editable surface + metric + time budget) with fuzziness audit and gate decision
Auto-Improvement Readiness Audit	5-layer prerequisite scorecard with gap remediation plan
Meta-Task Architecture Design	Full split architecture: meta-agent spec, task-agent spec, model-empathy constraint, handoff protocol
Trace Infrastructure Blueprint	Reasoning-trajectory logging schema, storage, analysis layer for any agent system
Emergent Behavior Playbook	Explicit harness affordances for 7 documented emergent patterns
4-Mode Safety Audit	Detection + response mechanisms for gaming, drift, contamination, cascade
Local Hard Takeoff Deployment Plan	End-to-end rollout: system selection → infrastructure → first 3 cycles → promotion → scale
Phase 2 Karpathy Audit	META — applies this skill to Antigravity's Phase 2 system; prescribes concrete upgrades (advisory only)

Workflow Table (by Tier)

Tier 1 — Foundation

#	Workflow	Slash	When Deployed
01	Karpathy Triplet Design	/nate-auto-triplet	Before any auto-improvement project — the gate
02	Auto-Improvement Readiness Audit	/nate-auto-audit	Assessing whether team/system is ready
03	Meta-Task Architecture Design	/nate-auto-architecture	Designing the split from scratch

Tier 2 — Practitioner

#	Workflow	Slash	When Deployed
04	Trace Infrastructure Blueprint	/nate-auto-traces	Before optimization infrastructure — foundation first
05	Emergent Behavior Playbook	/nate-auto-emergent	Designing/retrofitting a harness
06	4-Mode Safety Audit	/nate-auto-safety	Pre-production for any auto-improvement system

Tier 3 — Stacking + Meta

#	Workflow	Slash	When Deployed
07	Local Hard Takeoff Deployment Plan	/nate-auto-takeoff	Full project rollout (team or client consulting)
08	Phase 2 Karpathy Audit	/nate-auto-phase2	Leveling up Antigravity's own Phase 2 (advisory, no execution)

When to Deploy

• Building any self-improving or auto-optimizing agent system
• Evaluating whether a team/org is ready for auto-improvement
• Auditing existing Phase 2 / evolution / auto-research systems against frontier patterns
• Client consulting on deployment strategy for agentic auto-improvement
• Detecting metric gaming, silent degradation, contamination, or compounding errors
• Upgrading score-only logging to reasoning-trajectory logging
• Planning the 5-layer prerequisite build (context → trace → eval → sandbox → governance)

When NOT to Deploy

• Pure agent coordination problems (use nate-b-jones-orchestration-intelligence instead)
• Agent specification / containment (use nate-b-jones-agent-deployment-strategy)
• Intent capture and disambiguation (use nate-b-jones-intent-engineering)
• Systems with <3 agent runs (not enough history to optimize against)

Stacking Guide

Pair With	Combined Output	Use Case
nate-b-jones-orchestration-intelligence	Self-optimizing DPVI system	Multi-agent orchestration that improves coordination over time
nate-b-jones-trust-architecture	Autonomous-system red team	Safety audit for self-improving agents (metric gaming as attack surface)
nate-b-jones-intent-engineering	Intent-aware auto-improvement	Program.md becomes intent document
nick-saraev-bottleneck-thinking	Constraint-first triplet selection	Find the ONE editable surface worth optimizing
lance-yichao-context-engineering	Context-sufficient traces	High-signal traces preserving information meta-agent needs
Antigravity Phase 2 (native system)	Direct infrastructure upgrade via WF 08	Apply skill TO skill_benchmark.py + evolution loop

Key Principles

1. Constraints over intelligence — minimalism is the mechanism, not a limitation
2. Traces over scores — you cannot optimize what you cannot interpret
3. Prerequisites before loops — context, traces, eval, sandbox, governance or nothing
4. Earn the right — cheap-failure domains first, customer-facing last
5. Human judgment concentrates — framework design and program.md are higher-leverage than experiment execution
6. Same-model pairing — cross-model meta/task underperforms structurally
7. Iteration rate, not hit rate — 100 overnight beats 10 per day

Load Order

1. Load genius.md (required — all patterns, hidden knowledge, exemplars, signature moves, quality rubric, anti-patterns)
2. Load specific workflow for task
3. Reference files in references/ as needed per workflow
4. Cross-reference nate-b-jones-orchestration-intelligence/genius.md if designing self-optimizing orchestration