Here’s the uncomfortable truth the AI coding tool industry doesn’t want you to think about: the model barely matters anymore.
Every two weeks, a new LLM claims the coding crown. GPT-5.6 just dropped, and OpenAI is resetting limits to celebrate. Claude ships updates on a weekly cadence. Gemini jumps in and out of the conversation. The benchmark scores blur together. The winner changes by the time you finish reading the announcement.
And yet — the actual experience of using these tools hasn’t converged at all. Claude Code feels completely different from Copilot. Hermes works nothing like Codex. OpenCode is a world apart from Cursor.
Why? Because what you’re actually using is a harness — the software layer that sits between you and the model. And the harness is what determines everything about your experience: how it manages context, what tools it gives the model, how it handles multi-file edits, whether it can spawn sub-agents, how it integrates with your editor, terminal, and workflow.
The model inside is increasingly a commodity. The harness is the product.
The Commoditization Has Already Happened
Think back to 2023. The question “which AI should I use for coding” was genuinely meaningful. GPT-4 was dramatically better than anything else. Then Claude 3 Opus. Then Sonnet. Each leap changed what was possible.
That era is over. Today, the top 5 coding models are within spitting distance of each other on any realistic task. The gap between Sonnet 4.5 and GPT-5.6 on a real PR is dwarfed by the gap between a well-configured harness and a poorly-configured one. Swap the model behind a good harness and you’ll barely notice. Switch to a bad harness with the best model in the world and you’ll feel it immediately.
This is exactly what happened to databases. Twenty years ago, you’d agonize over Oracle vs. PostgreSQL vs. MySQL. Today, the query planner and developer experience of your ORM matters more than the underlying database engine for 90% of applications. The engine is a commodity. The abstraction layer is what you interact with.
Coding agents are going through the same commoditization — and most people haven’t noticed.
What a Great Harness Does
A harness isn’t just “a way to talk to an LLM.” A great harness provides:
-
Context orchestration: How does it pack your project structure, recent files, lint errors, and git history into the model’s context window? This is the hardest problem in coding agents, and every harness solves it differently.
-
Tool execution: Can it run tests, read files, grep your codebase, execute shell commands — and do so reliably without hallucinating flags?
-
Sub-agent management: Can it delegate tasks to parallel workers and merge results? Or is it strictly single-threaded?
-
Hook system: Can you inject custom behavior at every stage of the agent loop — before tool calls, after responses, on errors? This determines whether the tool fits your workflow or you fit the tool’s.
-
Plugin and MCP ecosystem: Is it a closed product or an extensible platform?
-
Failure recovery: When a tool call fails, does the harness retry intelligently, or does the whole task derail?
These are the dimensions that actually determine your productivity. Not whether the model is on version 4.5 or 5.6.
The Proof is in the Fumbles
The clearest evidence that harness design matters more than model quality? Watch the companies that get it wrong.
The recent “ChatGPT Work” rollout is a masterclass in misunderstanding the developer audience. Users who updated Codex found it replaced with a confusing dual-mode UI where chats and projects vanished, and “Work” and “Codex” modes appeared to do the same thing. The feedback was brutal — and deserved. The company optimized for product consolidation and brand unification, not for the developer experience of people who rely on a harness daily.
Meanwhile, the open-source harness ecosystem is thriving. Hermes adds MCP discovery and profile-local plugins. OpenCode ships with a robust hook system. Claude Code’s plugin architecture keeps getting deeper. These tools aren’t competing on model quality — they’re competing on how well they wrap the model.
This isn’t an accident. When your harness is open, extensible, and community-driven, it evolves faster than any product team can anticipate. When it’s a closed product subject to quarterly roadmaps and brand strategies, it stagnates — no matter what model powers it.
What This Means For You
If you’re still choosing your daily coding driver based on which model it supports, you’re asking the wrong question. Instead:
-
Evaluate the harness, not the model. Test how well the tool understands your project structure, handles your workflow, and surfaces context. The model can always be swapped.
-
Prioritize extensibility. A harness with a good hook system and MCP support will keep getting better. A closed harness will get worse relative to the ecosystem over time.
-
Watch for lock-in. If the tool makes it easy to switch models, that’s a feature, not a weakness. If you’re stuck with one provider’s model, you’re not using a harness — you’re renting one.
-
Build your own. The developer community is starting to realize that a coding agent is just a shell script with good UX around an LLM call. The barriers to building a custom harness are lower than ever.
The Future
In 12 months, we won’t argue about which coding model is best. That argument is already boring. We’ll argue about harness design: how to manage context, how to structure sub-agent delegation, what makes a great hook system, how to balance autonomy with safety.
The models will keep getting better, cheaper, and more interchangeable. The harness is where the real innovation — and the real value — lives.
The winners of the coding agent era won’t be the companies with the best model. They’ll be the ones that build the best harness.
Don’t confuse the two.