precision_manufacturing

Harness Engineering

The discipline of designing systems that make AI coding agents reliable — constraints, feedback loops, documentation, linting, review pipelines, and orchestration at scale
Co-Created by Kiran Shirol and Claude
Topics CLAUDE.md Constraints Review Pipelines Orchestration Entropy Management Agent Governance
home Learning Portal play_arrow Start Learning summarize Key Insights dictionary Glossary 8 chapters · 3 sections
Section 1

Foundations — Why the Harness Matters

What harness engineering is, why the model is commodity, and the evidence that the harness determines agent performance.
1
pets
What Is Harness Engineering?
The horse-tack metaphor. Formal definition: constrain, inform, verify, correct. Coined by Elvis Saravia in early 2026. Why the model is commodity but the harness is competitive advantage.
arrow_forward Learn
2
compare
The Model vs. The Harness
LangChain’s jump from 52.8% to 66.5% on Terminal Bench 2.0 by changing only the harness. Nate B Jones’s 78% vs 42% finding. AutoHarness from Google DeepMind (ICLR 2026).
arrow_forward Learn
3
description
Constraint Documents & CLAUDE.md
CLAUDE.md, AGENTS.md, .cursorrules files. The 3-tier architecture (root rules, skills, agent guides). Writing effective constraint documents that agents actually follow.
arrow_forward Learn
Section 2

Core Mechanics — Constraints, Review & Memory

Architectural enforcement, multi-agent review pipelines, feedback loops, and entropy management.
4
architecture
Architectural Constraints & Linting
Dependency layering enforced by CI. Deterministic linters and LLM-based auditors. “Vibecoded lints” targeting agent-specific mistakes. Pre-commit hooks and structural tests.
arrow_forward Learn
5
rate_review
Review Pipelines & Feedback Loops
Multi-agent review workflows. Self-verification loops and pre-completion checklists. Loop detection to prevent “doom loops.” The reasoning sandwich pattern.
arrow_forward Learn
6
memory
Memory & Entropy Management
Episodic memory for few-shot learning. Documentation consistency agents. Constraint violation scanners. Preventing codebase entropy drift over time.
arrow_forward Learn
Section 3

Scale & Future — Orchestration, Risks & Governance

Production orchestration systems, the four maturity levels, agent sprawl, and the future of the discipline.
7
account_tree
Orchestration at Scale
OpenAI Symphony (Elixir-based, dispatches Codex agents). Stripe Minions (1,000+ merged PRs/week). Basis ($200M revenue, zero human-written code). The four maturity levels.
arrow_forward Learn
8
shield
Risks, Governance & The Future
Agent sprawl (73% unmonitored “shadow agents”). Vendor lock-in at the workflow level. Security surface area. Building rippable harnesses. The future of the discipline.
arrow_forward Learn
link

Related Courses

layers
Context Engineering
8 chapters · Context management
code
AI-Assisted Coding
Coding with AI agents
smart_toy
Agentic AI
Agent architectures
security
AI Security
Security & safety