Building with AI - Harness Engineering
AI amplifies your codebase’s worst tendencies. The optional best practices that humans always cut under pressure become mandatory for AI.
Harness Engineering is the art of building self-improving constraint systems for AI agents. It’s not about making agents smarter, it’s about making the environment so well-designed that even a “dumb” agent can’t produce bad code. The harness becomes the codebase’s immune system, automatically detecting and correcting drift while agents generate at superhuman speed.
The New Non-Negotiables:
- 100% Code Coverage: Not for bug prevention, but to guarantee agents double-check every line they wrote. At 100%, the coverage report becomes a simple todo list with zero ambiguity.
- Thoughtful File Naming & Structure: Agents navigate via filesystem. ./billing/invoices/compute.ts communicates infinitely more than ./utils/helpers.ts.
- Fast, Ephemeral, Concurrent Dev Environments: Agents need short feedback loops (tests run in ~1 minute, not 30) and isolated environments (git worktrees, per-branch databases).
- End-to-End Types: TypeScript everywhere, OpenAPI-generated clients, Postgres types mapped to TypeScript. Types shrink the search space and eliminate illegal states.
- Small, Well-Scoped Files: Agents summarize/truncate large files. Small files load fully into context.
“Often, when teams struggle with agentic coding, it’s AI reflecting and amplifying their codebase’s worst tendencies.”
Harness = Context Engineering + Architectural Constraints + Garbage Collection
Planning (human-approved)
→ Test Generation (agent, TDD)
→ Implementation (agent, minimal code)
→ Adversarial Review (fresh context, zero tolerance)
→ Formal Verification (pure core)
→ Feedback Loop (update docs/linters/tests)
To learn more about Harness Engineering:
- The Emerging “Harness Engineering” Playbook → Explore how AI code generation shifts engineering toward harness building, verification systems, and orchestrating autonomous agents
- Harness Engineering → Learn a framework for harness engineering: building the infrastructure that makes AI-generated code production-ready and maintainable
- Harness engineering: leveraging Codex in an agent-first world → Understand how harness engineering enables AI agents to write production code by designing verification systems and guardrails
- AI Is Forcing Us To Write Good Code → Discover how AI code generation makes clean architecture and readable code essential since AI learns from your patterns
- Verified Spec-Driven Development → Explore a development methodology where AI agents generate formally verified code from specifications to guarantee correctness
- Les LLMs savent coder. Mais savent-ils maintenir ? → Découvrez pourquoi les LLMs excellent en code jetable mais peinent encore sur la maintenance long terme de projets complexes
Building with AI - Agents & Automation
- The 8 Levels of Agentic Engineering → Discover the progression from manual coding to fully autonomous AI agents across eight levels of engineering complexity
- Building Claude Code with Boris Cherny → See how Anthropic built Claude Code’s multi-agent system, achieving 90% success rates on complex engineering tasks
- Personal Software → Discover how personal AI agents will create custom software tailored to your unique needs and workflows
AI & People & Process
- How AI Will Reshape Public Opinion → Understand how AI-powered persuasion and personalized content will fundamentally alter how public opinion forms and spreads
- The looming AI clownpocalypse → Examine the risks of AI-generated code flooding the internet with low-quality content and eroding software reliability
- Will Features Even Exist? How AI Is Forcing SaaS To Rethink The Product Itself → Explore how AI transforms SaaS from feature factories to personalized experiences that adapt to each user’s unique needs
- Software Craftsmanship in the AI Era → Examine how software craftsmanship principles remain essential as AI shifts our role from code writers to architects and reviewers
- Coding After Coders: The End of Computer Programming as We Know It → Explore how AI coding assistants are transforming software development from writing code to directing intelligent agents
- How to Kill the Code Review → Examine how AI assistants are transforming code review from manual inspection to automated quality gates and verification systems
- My AI Adoption Journey → Learn Mitchell Hashimoto’s five-stage journey from AI skeptic to builder, including practical steps for effective AI adoption
Security
- How We Hacked McKinsey’s AI Platform → Learn how researchers exploited prompt injection vulnerabilities in McKinsey’s Lilli AI platform to extract sensitive documents
- We are building data breach machines and nobody cares → Examine how modern software systems are designed with security vulnerabilities by default, making breaches inevitable
- Anthropic won’t kill cyber, but it may kill some companies → Examine how AI-powered security tools from Anthropic will disrupt cybersecurity vendors but not eliminate the need for security
- How to steal npm publish tokens by opening GitHub issues → Learn how attackers exploited GitHub Actions workflows to steal npm publish tokens through malicious issue creation
- TruffleHog → Find and verify leaked credentials across Git repos, Docker images, filesystems with 800+ detectors and API validation
WebCraft, Architecture & Human Skills
The Illusion of Building → Understand why chasing trends and tools creates an illusion of progress while mastering fundamentals builds real lasting value
The two kinds of error → Understand the critical difference between recoverable errors that users can fix and unrecoverable bugs that developers must address
Stop Using Pull Requests → Explore alternative development workflows that replace pull requests with trunk-based development and continuous integration
The purpose of Continuous Integration is to fail → Understand why CI systems should fail fast and often, catching integration issues before they reach production
The Middle Loop → Learn about the middle loop between writing and deploying code: testing, debugging, and iteration often forgotten in developer tools
“All Lawful Use”: Much More Than You Wanted To Know → Examine the legal and ethical implications of AI training data and the “all lawful use” licensing framework
Tooling & Ecosystem
- Orbital data centers, part 1 There’s no way this is economically viable, right? → Explore the economics and technical challenges of building data centers in orbit for energy and cooling advantages
- backnotprop/plannotator: Annotate and review coding agent plans and code diffs visually, share with your team, send feedback to agents with one click. → Review and annotate AI agent plans visually, share with your team, and send feedback with one click
- Temporal: The 9-Year Journey to Fix Time in JavaScript → Discover why JavaScript’s Date API is fundamentally broken and how Temporal’s 9-year journey will finally fix time handling
- The Pulse: Cloudflare rewrites Next.js as AI rewrites commercial open source → Learn why Cloudflare rewrote Next.js in Rust and how AI is disrupting commercial open source sustainability models
Architecture
- Les tests d’architecture logicielle : 4 exemples pour les adopter → Découvrez 4 exemples pratiques de tests d’architecture pour garantir la conformité de votre code aux règles de conception
- Algebraic Effects: From Callbacks and Promises to Effect Handlers → Explore algebraic effects as the next evolution in async programming beyond callbacks, promises, and async/await