Same model, different tool structure, tripled the score.
Every AI agent has two parts. One is the model: the language model that reasons, generates, and predicts. The other is the harness: the environment, tools, memory, context, guardrails, and permissions that surround the model and tell it what to do, when to act, and what it can touch.
The common story treats the model as the agent. A smarter model produces a better agent. Upgrade the weights, and the agent improves. This story is wrong by a factor of two.
The harness and the model have been chasing each other for four years. When the harness asks for more than the model can deliver, agents fail. When the model catches up to the harness's expectations, agents produce breakthroughs. The gap between them is the single diagnostic variable of agentic AI history, as described in a framework from Dan McAteer, 2026. Understanding that gap explains every major agent failure and every major agent success since October 2022.
Picture two lines on a graph. The x-axis is time (2022 to 2026). The y-axis is capability.
The first line is harness ambition: what the surrounding environment asks the model to do. Browse the web. Write code across files. Run multi-step tasks without human input. Manage memory. Select tools. The harness defines the scope of autonomy the agent attempts.
The second line is model capability: what the model can reliably deliver under those conditions. Reason over long contexts. Follow complex instructions. Recover from errors. Maintain coherence across turns.
When harness ambition sprints ahead of model capability, agents break in public. When model capability catches up or surpasses the harness, breakthroughs follow. The gap between these two curves is not a side effect. It is the story.
Capability ^ | Harness Ambition | / \ Model Capability | / GAP (failure) \ / | / \ / GAP migrates | / \/ to human boundary | / / \ | / Model Capability / \ Harness Ambition | / / \ |/________________________/ \_______________ +---------------------------------------------------> Time 2022 2023 2024 2025 2026 AutoGPT Cursor o1 Claude Code (fail) (correct) (leap) (breakout)
The harness-model gap is the diagnostic variable. Misaligned curves cause failures. Converging curves produce breakthroughs.
The curve-gap model is visible in four years of agentic AI history. Every milestone either shows the curves misaligned or converging, and the outcome follows from the gap position.
In October 2022, the ReAct paper introduced a foundational idea: give a language model both reasoning and action capabilities inside the same loop. The harness was minimal. The model was early. The curves sat close together, and the concept worked in constrained settings.
By February 2023, Toolformer showed models could learn to call external APIs on their own. The harness stayed modest. The gap stayed small. Progress was real but quiet.
Spring 2023 broke the pattern. AutoGPT and BabyAGI arrived with harnesses that assumed near-infinite autonomy: browse, write, plan, execute, self-correct, all without human input. The harness ambition curve spiked. The model underneath (GPT-3.5 and early GPT-4) could not deliver on those promises. Agents looped, hallucinated, and burned API credits on tasks they could not complete. The gap was the widest it had been, and the product failed in the most public way possible.
The 2023-2024 correction came from the other direction. Cursor and GitHub Copilot narrowed the harness. Instead of full autonomy, they scoped agents to code completion within a file, a repository, a defined context window. The harness stopped asking for more than the model could give. The curves aligned, and developers adopted these tools by the millions.
Late 2024 brought a model-side leap. The o1 reasoning model expanded what the model curve could deliver: longer chains of thought, better error recovery, stronger multi-step planning. Boris Cherny's team at Anthropic started building harnesses that assumed this new capability, preparing for the overhang. Lukasz Kaiser later noted a performance jump over Christmas 2025 when reasoning models crossed a threshold.
Claude Code launched in February 2025. By May 2025 it reached general availability. The growth from there is the clearest proof of curve alignment in the history of AI agents: $1B ARR by November 2025, $2.5B ARR by February 2026, $8B ARR by May 2026, 54% market share in AI coding tools, and 4% of all GitHub commits. SemiAnalysis projects that figure reaches 20% by end of 2026. Claude Code's ARR now exceeds Anthropic's entire company revenue from early 2025.
| Date | Milestone | Harness | Model | Gap | Outcome | |------|-----------|---------|-------|-----|----------| | Oct 2022 | ReAct paper | Minimal (reason + act loop) | Early LLMs | Small | Concept proven | | Feb 2023 | Toolformer | Modest (API calling) | GPT-3.5 era | Small | Quiet progress | | Spring 2023 | AutoGPT/BabyAGI | Full autonomy assumed | GPT-3.5/4 | Massive (harness ahead) | Public failure | | 2023-2024 | Cursor/Copilot | Scoped to code context | GPT-4 | Narrow (intentional) | Mass adoption | | Late 2024 | o1 reasoning model | Existing harnesses | Reasoning leap | Model catches up | Overhang builds | | Feb 2025 | Claude Code launch | Matched to Sonnet 3.5 | Strong coding model | Aligned | Product-market fit | | May 2025 | Claude Code GA | Expanded autonomy | Model supports it | Aligned | Growth accelerates | | Nov 2025 | $1B ARR | Broadened tool access | Model improved | Tight | Revenue milestone | | Feb 2026 | $2.5B ARR | Wider deployment | Claude 3.5+ | Tight | Market leadership | | May 2026 | $8B ARR, 54% share | Full coding harness | Claude Sonnet 4+ | Aligned | Category dominance |
Four years of agentic AI mapped as harness-model alignment. Every failure and breakthrough correlates to gap position.
Two independent benchmarks from 2026 put numbers on the curve-gap claim.
Harness-Bench tested 106 realistic end-to-end tasks across 5,194 trajectories, pairing 6 different harnesses with 8 model backends. The result: swapping the harness while keeping the model constant produced larger performance deltas than swapping the model while keeping the harness constant. NanoBot scored 76.2% and OpenClaw scored 52.4% on the same model. A 23.8-point spread, driven by harness design alone.
The ARC-AGI-3 results from OpenAI tell the same story from a different angle. GPT-5.6 Sol without retained reasoning or compaction scored 7.8%. With only the official harness settings, it scored 13.3%. With both retained reasoning and compaction enabled, it scored 38.3%, a tripling, at roughly 6x fewer output tokens. The model was identical in all three runs. The harness settings changed. Fran\c{c}ois Chollet noted the distinction between general-purpose API settings and benchmark-specific custom harnesses, which matters for how you interpret the numbers. But the core arithmetic stands: same model, different harness, triple the score at a fraction of the compute.
| Benchmark | Harness/Settings Changed | Model Unchanged | Performance Delta | |-----------|-------------------------|-----------------|-------------------| | Harness-Bench | NanoBot vs. OpenClaw | Same model backend | 23.8-point spread (76.2% vs. 52.4%) | | ARC-AGI-3 | None vs. official vs. retained reasoning + compaction | GPT-5.6 Sol | 7.8% to 13.3% to 38.3% (tripling at 6x fewer tokens) |
Harness changes produce larger performance deltas than model swaps. Capability is co-engineered, not model-centric.
The implication for anyone building agents is direct. Spending six months waiting for a better model while running a mediocre harness is the wrong trade. The harness is where half the score comes from. It is also where the builder has control.
The harness is not permanent. It is scaffolding.
Capabilities start outside the model, in the harness: tool selection logic, memory management, multi-step planning, error recovery routines. Over time, models absorb these capabilities through training. Once absorbed, the external scaffold becomes deletable. The pace of agentic progress is measured by how much harness you can remove while the agent retains its performance.
The clearest proof of this cycle completing is Claude Code's system prompt. Anthropic's Thariq Shihipar confirmed that 80% of Claude Code's system prompt was deleted when the newest Claude 5 (Fable) generation arrived, with no loss on internal benchmarks. The discussion at AI Engineer's World's Fair expanded on why: instructions that once told the model how to behave were now baked into the model's own weights. Keeping those instructions in the prompt was not neutral. For Fable-class models, capability instructions cap performance because they conflict with internalized behavior. The control surface becomes a context surface. Hard rules move from prose to code, implemented as hooks rather than prompt instructions.
GPT-5.1 Codex-Max shows the same cycle on the model side. It is the first model natively trained to operate across multiple context windows through compaction, a capability that previously required external harness engineering. What was harness is now model.
The next candidates for absorption are visible: multi-agent orchestration, tool selection, long-term memory. Self-improving harnesses are under active research. The cycle will continue. Train the model on harness behavior. The model absorbs it. Shed the harness component. Build the next layer of harness on top.
Here is where the two-curve model separates from a simpler "models are eating software" take.
When the model absorbs computer-facing harness capabilities (tool selection, memory, planning), the gap between harness ambition and model capability does close on that axis. But it does not vanish. It reconstitutes at the human boundary.
The old gap was between what the harness asked of the model and what the model could deliver. The new gap is between what the agent asks of the human and what the human can provide. As Ryan Lopopolo put it: "The only fundamentally scarce thing is the synchronous human attention of my team."
Tokens are abundant. Compute is scaling. Model capability is growing quarter over quarter. Synchronous human attention is not growing at all. It is fixed. A senior engineer has the same number of waking hours as she did in 2022. The gap reconstitutes at the point of maximum scarcity.
This produces an inversion. The harness was born as the human's interface to the model: "here is how I want you to work." It inverts into the model's interface to the human: "here is what I need from you, and here is how I will ask." Anthropic's long-running agent progress files and agentic approval queues are the first visible examples of this inversion. The agent produces work, then asks for human attention at specific, defined moments.
McAteer predicts that within a year, every agentic AI company will ship a human attention policy surface, analogous to AGENTS.md files. Where AGENTS.md tells the agent how to work with your codebase, the attention-interface tells the agent how to work with you: which decisions require approval, when to interrupt, what to batch. And like everything else in agentic systems, it will become learnable. Every correction becomes training data for better interruption timing.
The gap between permissions the agent needs and attention the human can give is not a bug. It is a structural feature. Some decisions require human judgment because they involve identity, trust, and organizational legibility, not because the model lacks capability. No amount of model improvement absorbs the question "should we ship this to our largest customer tomorrow?" That requires a human who owns the outcome.
The two-curve framework is a diagnostic tool you can apply to your own system today.
If your harness assumes capabilities your model cannot deliver, you are repeating AutoGPT's mistake. Your agents will loop, hallucinate, and burn resources on tasks they cannot complete. The fix is not a better model. It is a narrower harness, matched to what the model can do now.
If your model has outgrown your harness, you are leaving Claude Code-scale value on the table. The 23.8-point Harness-Bench spread and the 13.3% to 38.3% ARC-AGI-3 arithmetic prove that harness engineering is not a secondary concern. It is half the system.
If your harness still carries capability instructions that a Fable-class model has internalized, those instructions are capping your agent's performance. Delete them. Test the benchmarks. The model absorbed them months ago.
The gap is the business opportunity. It always has been. The question is which side of the gap you are building on, and whether you have noticed that the gap has moved. It no longer sits between your harness and your model. It sits between your agent and your team's attention. The companies that build the attention-interface, the system that governs how agents request and respect human attention, will own the next layer of agentic value. The rest will keep optimizing a gap that has already closed.