The optimization mentioned can be done more effectively than just switching the model, and it relates to a harness.<\/p>\n\n\n\n
In the context of coding agents, a harness is the layer around the model that determines how the model operates:<\/p>\n\n\n\n
- \n
- what context it receives,<\/li>\n\n\n\n
- which tools it can use,<\/li>\n\n\n\n
- how it interacts with the codebase,<\/li>\n\n\n\n
- how it executes and verifies changes,<\/li>\n\n\n\n
- and how the overall task is orchestrated.<\/li>\n<\/ul>\n\n\n\n
A 2026 survey of agent systems (arXiv:2606.20683<\/a>) compares the same Claude models across different harnesses on a well-established benchmark (SWE-bench Verified). The numbers jump from 33.6% to 49.0% of tasks solved, approximately a 46% relative gain when only the scaffolding around the model changes. Even at the frontier, harness choice still moves the score from 76.8% to 80.9%.
And this idea is starting to show up beyond software engineering benchmarks. In August 2026, NVIDIA demonstrated just how much the surrounding system can matter on ARC-AGI-3: the same Claude Opus 5 reportedly went from 30% to 100% after being placed inside a more sophisticated harness, including better memory management and a supervisory component.<\/p>\n\n\n\nAs TechCrunch put it,<\/p>\n\n\n\n
\n
The harness \u2013 rather than the model itself \u2013 is increasingly becoming the real hero of long-horizon agentic tasks<\/em><\/em>.<\/p>\n\n\n\n
TechCrunch, 21.08.2026<\/a><\/em><\/p>\n<\/blockquote>\n\n\n\n
What we’ve learned in production is now confirmed by the research community. The model \u2013 this legendary subject of the very first “what do you develop with” questions is a commodity everyone can rent at the same prices, through the same APIs.<\/p>\n\n\n\n
Yet with the same model, we’ve seen results ranging from statistical noise to a tenfold acceleration. Having investigated, we found the harness to be far more important, and we treat it as the engineering environment you build around your project so that it works within your SDLC, not beside it.<\/strong><\/p>\n\n\n\n
The evidence keeps piling up<\/strong><\/h2>\n\n\n\n
In our recent analysis of the two newest flagship models, Beyond the flagship: what Fable 5.1 and GPT-6 Astra tell us about AI-native delivery<\/a>, we went through everything measurable published in their first week.<\/p>\n\n\n\n
The most durable findings were not about either model, but about the systems around them:<\/p>\n\n\n\n
- \n
- Same weights, different scaffolding. ARC Prize reported<\/a> GPT-6 Astra at 62.7% on ARC-AGI-3 under a provider-neutral harness and 98.6% with OpenAI’s own Provider Adapter. The difference is that the adapter keeps the model’s reasoning state between actions and manages the context itself.<\/li>\n\n\n\n
- Benchmarks increasingly measure systems. Real-SWE<\/a> from Specific Labs, built on private production repositories, scores model-and-harness pairs rather than models: Fable 5.1 in Claude Code leads at 38.8%, and Astra in Codex CLI reaches 33.8%, with no configuration above 40%. The most common reason for failure is a missed requirement, exactly the kind of gap that project-specific context and verification steps are meant to close.<\/li>\n<\/ul>\n\n\n\n
This is the nuance behind our title. It is not enough to pick a harness, just as it is not enough to pick a model<\/strong>. What matters is how well the harness configuration fits your codebase, your workload, and your delivery process, and whether you measure it on your own tasks rather than on a public leaderboard.<\/p>\n\n\n\n
We were our own Client Zero<\/strong><\/strong><\/h2>\n\n\n\n
Such research is time-consuming, which is why Sii Poland comes with a good part of it done for you already. We simply tested it on ourselves.<\/p>\n\n\n\n
Our internal delivery organization (more than 115 applications and platforms and over 280 integration modules across .NET, Angular, Microsoft Dynamics 365, WordPress, and heavyweight integration middleware) became Client Zero.<\/p>\n\n\n\n
Every AI-augmented team has access to the same models. What we engineered was the harness:<\/p>\n\n\n\n
- \n
- organizational developer kits with project-specific context rules,<\/li>\n\n\n\n
- specialized agents for implementation, review, testing, and bug reporting,<\/li>\n\n\n\n
- MCP servers wired into real infrastructure,<\/li>\n\n\n\n
- and safeguards that block secrets and require developer review.<\/li>\n<\/ul>\n\n\n\n
- Benchmarks increasingly measure systems. Real-SWE<\/a> from Specific Labs, built on private production repositories, scores model-and-harness pairs rather than models: Fable 5.1 in Claude Code leads at 38.8%, and Astra in Codex CLI reaches 33.8%, with no configuration above 40%. The most common reason for failure is a missed requirement, exactly the kind of gap that project-specific context and verification steps are meant to close.<\/li>\n<\/ul>\n\n\n\n
- Same weights, different scaffolding. ARC Prize reported<\/a> GPT-6 Astra at 62.7% on ARC-AGI-3 under a provider-neutral harness and 98.6% with OpenAI’s own Provider Adapter. The difference is that the adapter keeps the model’s reasoning state between actions and manages the context itself.<\/li>\n\n\n\n
![\"\"](\"https:\/\/sii.pl\/blog\/wp-content\/uploads\/2026\/04\/Banner-AI-Offer-Sample-Desktop.jpg\")
\n <\/picture>\n