Why Everyone in AI is Suddenly Obsessed with Harnesses

A frontier model today rarely works alone. It reasons with tools, retrieves information, remembers context, and relies on an orchestration layer that increasingly defines its performance.

Memory, orchestration, evaluation, tool use, security, observability, prompt management, and workflow automation are becoming the new competitive edge. 

In short, the value is shifting from the model to the harness built around it.

Some of the most widely used agent harnesses today include Anthropic’s Claude Code, OpenAI’s Codex, Google's Jules, OpenHands, Aider, Cline, Devin, and OpenCode.

Whether it’s an MCP server, an agent framework, a memory layer, an evaluation platform, or a tool-calling SDK, the industry’s centre of gravity now lies in the surrounding stack built around the model.

The Industry Weighs In

The industry’s growing obsession with harnesses is becoming hard to ignore. 

Sarvam AI co-founder Pratyush Kumar tells AIM that the biggest gains in AI will come from combining powerful models with domain-specific harnesses rather than relying on models alone. 

“The outcome that you get from a model plus harness is surprisingly much more complex compared to just hitting the model,” he says. “In fact, it will seem strange in one or two years to hit the model. It will feel like you’re doing something stupid.”

Kumar explains that harnesses can be customised for domains such as Android development, data science, and cybersecurity while sharing the same underlying foundation. “What harnesses really do is bring domain expertise to models,” he adds.

That view is shared by Xiaoyin Qu, Founder of AI startup Tycoon.us, who argued that even the strongest models struggle to deliver value without the right software layer around them.

“Harness is so, so important. DeepSeek V4 Flash is a great model, but without a great harness like Codex or Claude Code, people can’t seem to get enough value out of it,” she wrote on X. “My take is agents are not smart enough. Model is….There are still lots of improvements on the harness side that make agents more approachable.” 

Comparing today’s AI agents to a box of Lego bricks, Qu said most users are handed “1,000 Lego bricks with no manual” when what they really need is a guided kit that helps them build something useful.

Models are Becoming Interchangeable

During the recent earnings call, Microsoft CEO Satya Nadella said the company is building a new model system where the harness, context, memory, and action space are separate from any one model family. He added that this moves the frontier on the cost-to-outcome curve. 

“It also has the added benefit of business continuity and resilience because every model is substitutable,” Nadella said.

In other words, Microsoft wants enterprises to achieve better business outcomes at lower cost by treating foundation models as interchangeable components rather than the centre of the AI system.

Muratcan Koylan, Member of Technical Staff, Research at healthcare AI company Sully.ai, believes that harnesses should be designed with future models in mind rather than today’s capabilities. “Anticipate where they’ll be in six months, and build your architecture for that trajectory,” he wrote on X.

While many conversational capabilities are gradually being absorbed into models, the orchestration layer may be indispensable. 

"The real-time front can own timing, interruptions and conversational repair. But for tasks that require tool calls or deeper reasoning, you still need a harness," he said. "The conversational harness is moving into the model; the transactional harness is not."

Elvis Saravia, Co-founder of AI research and education community DAIR.AI, believes that choosing the right agent harness has become a critical skill for AI engineers. Research published on arXiv finds that the same model, task, and prompt can produce a 5–30x difference in cost per successful outcome simply by switching between agent harnesses. 

“Harness design and prompt wording decide most agent spend before the model reasons at all, and both are cheap to change,” he wrote on X.

The idea that the harness matters more than the model is already resonating with developers. In a recent blog post, Burke Holland, Principal Developer Advocate at Microsoft for GitHub, argued that “the biggest gains in my productivity” come not from installing every new AI tool or mastering prompt tricks, but from “how I use the harness and how well I understand it.”

He described GitHub Copilot as an agent harness that handles planning, orchestration, execution, and review, allowing developers to focus on solving problems rather than managing AI workflows. “The harness is all you need—mostly,” he wrote, arguing that consistent workflows matter more than constantly chasing the latest models, MCPs, or custom agents.

However, Adithya S Kolavi, Research Engineer at Hugging Face, believes the debate is not about choosing between models and harnesses, but understanding where each creates value. “For companies that have the budget and capability to train or fine-tune their own models, there's still a lot of value to be created at the model layer,” he tells AIM. 

“That said, harness engineering has become incredibly important as models have become more capable,” he notes.

Kolavi adds that the choice ultimately depends on the company’s goals. “I don't think it’s an either-or situation. Both the model and the harness matter, and where you invest depends on your capabilities, compute budget, and the end solution you're trying to build.”

Harness or Model? 

Abhishek Harshvardhan Mishra, Founder of AI research firm Deus Experiments, observes that the industry has entered an era of model-harness co-design. “Almost all the labs are using a harness to train their models on all kinds of tasks such that their own harness can always do better than any other alternatives," he remarks.

Mishra adds that models perform significantly better when they are deeply familiar with their operating environment. However, he believes the biggest returns today come from improving the software around the model. “In terms of pure ROI, improving harness experience has the best returns.”

He believes harnesses complement models rather than replace them. 

“In terms of capability, a harness cannot bring out from a model what it isn't capable of,” he says. “But it can reduce variance by a lot and follow good practices in steering and context management to bring out its best.”

He points to edge AI coding agent Oh My Pi as an example of how a well-designed harness can consistently improve the performance of existing models without changing the underlying model itself.

Meanwhile, Kolavi believes the biggest difference between a good harness and a great one lies in efficiency rather than raw capability. “Assuming you already have a strong underlying model, a well-designed harness can reach the same solution with fewer tokens, lower latency, and better reliability through better orchestration, tool use, and planning,” he says.

At the same time, harnesses cannot compensate for a weak foundation model. "A harness is only as good as the model driving it. If the underlying model lacks the reasoning ability or capabilities needed for the task, there's only so much a harness can do,” Kolavi says.

Kolavi believes there is no one-size-fits-all approach to deciding whether to invest in models or harnesses. “It depends on the compute budget, your team’s capabilities, and the problem you’re trying to solve,” he says.

For most startups, he sees the software layer as the more practical investment. “It lets you move quickly, iterate on the product, and focus on solving user problems instead of spending significant resources training models from scratch,” Kolavi explains.

However, he opines that organisations with the compute resources and expertise to train or fine-tune their own models can still gain meaningful advantages by investing in the model layer.