The Coming Loop: How AI Harness Engineering Is Quietly Rewriting Software Development
By TekMag Editorial Team — June 24, 2026
Core Thesis: The Loop is the Product
The most important shift in AI-assisted coding has nothing to do with better models. It is about the loop.
"I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops." — Boris Cherny (PyTorch early tooling)
This defines Harness Engineering: the third and most consequential phase of AI-assisted development. The implications for how software teams structure work, review code, and think about productivity are only beginning to surface.
The Equation: Agent = Model + Harness
"A decent model with a great harness beats a great model with a bad harness." — Addy Osmani (Google)
The Harness (as defined by Birgitta Böckeler of Thoughtworks on Martin Fowler's blog) is everything in an AI agent except the model itself:
- System prompts, tools, MCP servers
- File system access and execution sandboxes
- Verification hooks and orchestration logic
- The scaffolding that turns a stateless LLM into something that can actually finish a task
Empirical proof: LangChain's Vivek Trivedy moved his team from 30th to 5th place on Terminal Bench 2.0 by changing only the harness, keeping the same underlying model untouched.
The Two Loops (Armin Ronacher)
Most people conflate the loops. They are fundamentally distinct:
Inner Loop (Agent Loop)
The model calls tools, incorporates results, edits files, runs tests. This is what most current AI coding tools do — a self-contained cycle of act-observe-iterate.
Outer Loop (Harness Loop)
An external orchestration layer that queues tasks, attempts them, evaluates whether work is genuinely done, then injects fresh context, starts a new session, or hands off. The task outlives what the model would normally call "done."
Why it matters: Most teams only build the inner loop. The magic — and the risk — lives exclusively in the outer one. Dedicated orchestration platforms like EverOS are emerging as commercial harness infrastructure precisely because the outer loop requires different design thinking than the inner loop.
The Three Phases of AI Engineering (Neely Dunlap / Faros AI)
| Phase | Era | Focus | Output |
|---|---|---|---|
| 1 — Prompt Engineering | 2022-23 | Getting right output from a single prompt | Snippets and boilerplate |
| 2 — Context Engineering | 2024-25 | Managing relevance and memory across conversations | Features and file updates, needs human babysitting |
| 3 — Harness Engineering | 2026 | Building autonomy and control systems around models | End-to-end tasks, entirely new failure modes |
The Trigger: Models became capable enough that the bottleneck shifted from "can the AI write code" to "can the human write a spec the AI can execute against."
"I am bottlenecked on specs. The agent loop is less of a thing for me now... The gate is almost entirely on spec creation and PR review." — Senior engineer, Hacker News discussion (318 points, 226+ comments)
Where Loops Win
- Porting / Code Transformation: Ronacher's team used Claude to port MiniJinja from Rust to Go. Boring, mechanical work that loops handle perfectly.
- Performance Exploration / Benchmarking: Setting up and iterating through benchmark configurations.
- Security Scanning: The curl project's "summer of bliss" (a month off from vulnerability reports) is cited as a canary in the coal mine — security researchers are already looping against open source at scale.
- Research / Proofs of Concept: Artifacts with no long shelf life where speed matters more than maintainability.
Where Loops Struggle
- Production Code Needing Deep Human Understanding: Loops amplify the model's worst tendencies — too defensive, too complex, too local in reasoning.
- Systems Needing Strong Invariants: Models avoid hard constraints, adding fallbacks instead of making bad states impossible. > "If each iteration adds another small defense, the system slowly becomes less understandable while appearing more robust." — Armin Ronacher
- Code Where Taste, Judgment and Maintainability Matter: Ronacher reports Claude Code with ultracode (30+ minute uninterrupted runs) produces worse code than autumn 2025 with more human involvement.
The Dark Side: Comprehension Debt
Comprehension Debt (coined by Addy Osmani): The speed of automated code generation outpaces the team's capacity to review it. Codebases emerge where underlying design decisions are completely unmapped.
The Two Hostile Camps
| Camp | Behavior |
|---|---|
| "Agent Go Brrr" | Loops on loops, massive PRs, cannot explain their own code — but shipping fast |
| Exhausted Reviewers | Inundated with AI-generated PRs, burning out |
"Reviewing AI slop is data labelling... You used to be an F1 driver. Now you're an instructor for bad drivers." — Hacker News commenter
The Statistics
- 30% of developers report little to no trust in AI-generated code (DORA report via Augment Code)
- 10,000+ new security findings per month introduced by AI-generated code by September 2025 (Apiiro) — a 10x increase from December 2024
- Spotify Honk: Merged 1,500+ AI-generated PRs across hundreds of repos since mid-2024
- Five-person teams doing the work of fifty
Why This Is Permanent
Three forces make harness engineering a structural inevitability:
- Security: Attackers are already looping. Defenders must loop too, or lose.
- Economics: Five people doing the work of fifty is not a bug — it is the endgame.
- Technical Debt: Codebases that resist AI intervention (strong invariants, deep architectural reasoning, taste-dependent design) will become increasingly expensive to maintain relative to AI-compatible codebases.
"Your job is to write loops." — Boris Cherny
Originally published on TekMag.
Top comments (0)