Schema: Frontier Models with the Right Harness Achieve ~99% on ARC-AGI-3 Public

I'm Matt Wood, and this is For Your Information. A live list of riffs and links for you and your agent, drawn from what I'm reading, noticing, questioning, concluding, and revising.

Connections

Supports

Kiro CLI Custom Agents The 'right harness' thesis for frontier models directly explains why Kiro's custom agent configuration matters as much as the underlying model.
The Last Harness Youll Ever Build arXiv ARC-AGI-3's finding that harness quality (not just model capability) drives near-99% performance backs the paper's thesis that a well-built harness is the durable investment.

Related to

Harness engineering: leveraging Codex in an agent-first world | OpenAI Both discuss how the surrounding harness/environment interface, not raw model power, determines agent performance on complex tasks.

Supported by

DeepSeek Harness: Plugin Architecture Schema's finding that frontier models with the right harness achieve ~99% on ARC-AGI-3 directly validates the premise that plugin-based harness architecture significantly amplifies model capability.
The Shape of Things to Come, Part 1: The Continuous Thunderdome Yegge's argument that loops/graphs harnesses unlock autonomous large-scale problem solving directly supports Schema's finding that the right harness—not just the model—is what achieves near-perfect performance on hard benchmarks
Building an Advanced Agentic Harness The harness architecture described (typed validation, verification hierarchies, dependency graphs) directly explains the mechanism by which frontier models achieve ~99% on ARC-AGI-3 — structured scaffolding compensates for raw model limitations
Improving Gpt 5 6 Sol In Chatgpt OpenAI's model improvements on problem-solving benchmarks directly supports the finding that frontier models with the right harness can achieve ~99% on ARC-AGI-3, as base model capability improvements compound with agentic enhancements
Pi, Minimal and Performant Pi's benchmark-leading performance with minimal tooling supports Schema's finding that the right harness design—not just model capability—drives near-perfect task performance
Discovery Loop Both demonstrate that frontier AI models combined with the right harness/infrastructure can dramatically exceed baseline performance — Discovery Loop applies this principle to experimental science rather than benchmarks

Challenged by

LLMs can't jump Schema's claim that frontier models with the right harness achieve ~99% on ARC-AGI-3 is directly challenged by the thesis that LLMs lack manipulative abduction and physical simulation — benchmark success may not reflect genuine inventive or scientific reasoning capability.
What Sort of Maths Are LLMs Good At? Frontier models achieving ~99% on ARC-AGI-3 with the right harness suggests broader reasoning capability that the new item's analysis of mathematical limitations may need to account for or qualify

Grok 4.6 Benchmarks and Cost Efficiency Analysis Both examine frontier model performance benchmarks (Artificial Analysis Intelligence Index vs ARC-AGI-3), contributing to the same discourse on measuring AI capability
Bullet: Fast Coding Agent Both items focus on high-performance coding agents achieving near-frontier results on software engineering benchmarks (SWE-bench vs ARC-AGI-3), with architectural optimizations driving benchmark scores
Material Discovery Bench: LLM Research Benchmark Both are benchmark frameworks evaluating LLM capabilities on specialized scientific/technical tasks, with leaderboards tracking model performance on challenging discovery problems

Type:	link
First published:	Monday, 20 July 2026 at 12:00 UTC
Location:	https://schema-harness.github.io/
URL:	https://mattwood.fyi/i/c5a78c19