There's a war happening right now over where AI value actually lives. Not in headlines—in architecture decisions, pull requests, and enterprise file systems. Today's digest gets into all of it.

The Harness Wars: Does the Scaffold Matter?

A debate that's been simmering for months is finally forcing a real answer. The question: does the "harness"—the orchestration layer that wraps a model—actually create value? Or does it just get in the way?

The Big Model camp is blunt. Claude Code's Boris Cherny describes their approach as deliberately minimal. Quote: "All the secret sauce is in the model. This is the thinnest possible wrapper. We literally could not build anything more minimal." OpenAI's Noam Brown goes further—arguing that before reasoning models, complex scaffolding was necessary to coax reasoning behavior. Now? "You just give the reasoning model the same question without any scaffolding and it just does it." His implication: today's scaffolds will be replaced by tomorrow's models.

Here's what's interesting, though. The Big Harness camp has data too. Jerry Liu argues the harness is everything—that the biggest barrier to value is your ability to engineer context and workflow around a model. And a recent experiment showed dramatic improvements across fifteen different models just by optimizing the prompt harness. Same models. Better wrapper. Better results.

Swyx—who's been tracking this closely—notes the evidence is genuinely mixed. Scale AI's benchmark finds Claude performs two-and-a-half points better inside Claude Code versus a generic agent scaffold. But GPT shows the reverse. The harness effect is essentially noise within the margin of error.

Simon Willison adds a sharp practical angle here. He's documenting anti-patterns in agentic engineering—the behaviors that make harnesses fail. The biggest one: submitting unreviewed code to collaborators. If you dump a thousand lines of agent-generated code into a pull request without reviewing it yourself, you're not delivering work—you're delegating the actual work to someone else. A good agentic pull request has small, testable changes, human-verified functionality, and notes on how you tested it. The harness matters. So does the human using it.

Why Enterprise Agents Are Harder Than Coding Agents

Aaron Levie, CEO of Box, spent time this week explaining something most AI coverage glosses over: why coding agents achieved escape velocity while enterprise knowledge work agents haven't.

His enumeration is worth sitting with. Coding has almost every favorable condition imaginable. The entire codebase is accessible to any engineer. The medium is purely text in, text out. The models were trained heavily on code. The people building the AI tools are daily users of those tools—which means rapid feedback loops. And developers are technical enough to install whatever's newest.

Knowledge work has the opposite conditions. A banker might have access to a tiny slice of relevant data—locked behind permissions, in different formats, spread across organizations. There's no specification culture. No documentation standards. And crucially—Levie's sharpest point—when a coding agent produces slop, the end user often doesn't notice. When a knowledge work agent produces a contract with a slightly wrong clause, or a memo with inconsistent fonts, the end user notices instantly.

This leads to what Levie calls the context engineering problem—and it's severe. Imagine sixty thousand tokens of usable context window against a corpus of fifty million pages of enterprise data. That gap doesn't get bridged by bigger context windows alone. It requires models that know when to stop searching. That can smell when a document doesn't match a query. That can judge when they've exhausted the search space rather than just returning six out of ten answers with a shrug.

His concrete test case: ask an agent to find Box's ten office addresses. Lower-tier models return six addresses and say they couldn't find the rest. They don't know what they don't know. The newer Opus models—four-five, four-six—show real judgment here. They'll try multiple query strategies, recognize they're not converging, and stop rather than hallucinate.

Levie's prediction: twenty-twenty-five was the coding agent year. Twenty-twenty-six is the knowledge work agent year. But it'll require enterprises to fundamentally restructure how they store and permission their information—not because agents demand it, but because the productivity gap between companies that do this and those that don't will compound fast.

One more thread from this conversation that deserves attention: agent identity. As enterprises deploy agents that act on behalf of employees, they face a question nobody has fully solved. An agent isn't a user. You can't just give it an account. The creator of an agent probably inherits liability for what it does. You need oversight structures, access controls, and sandboxed workspaces—a kind of filesystem that's partitioned for agent use while remaining auditable by humans. Levie frames this as one of the genuinely hard infrastructure problems of the next two years. Boring to ninety-eight percent of the internet. Critical to the other two percent.

Qwen's Crisis: A Landmark Open Source Team in Turmoil

The most significant model news this week has nothing to do with a benchmark. The Qwen team at Alibaba—builders of arguably the best open-weight model family on the planet—appears to be fracturing.

Lead researcher Junyang Lin announced his resignation on X in a single sentence: "Me stepping down. Bye my beloved Qwen." Within hours, several core team members followed. Binyuan Hui, who led Qwen's code development and the entire agent training pipeline from pre-training through post-training. Bowen Yu, who led post-training research. Kaixin Li, a core contributor to Qwen's latest vision and coding models. Multiple younger researchers on the same day.

Simon Willison, who's been following open-weight model development closely, frames why this hits hard right now. The Qwen three-point-five family—released over the past few weeks—is exceptional. The flagship is a nearly eight-hundred-gigabyte mixture-of-experts model. But the smaller siblings are what's remarkable: a two-billion-parameter model that does reasoning and vision, compresses to under five gigabytes, and is—by most accounts—surprisingly capable for its size. The Qwen team had found a genuine methodology for extracting performance from smaller and smaller models. That's the capability that's now at risk.

Alibaba's CEO held an emergency all-hands. Lin posted a vague reassurance on WeChat: "Brothers of Qwen, continue as originally planned, no problem." Whether he's returning or just steadying the ship on his way out is unclear. The trigger appears to be a reorganization that placed a researcher hired from Google's Gemini team above Lin—though this detail hasn't been confirmed.

—

The thread connecting all of this: we're in a moment where the leverage is shifting. Models are getting better, which pressures the harness. Agents are getting more capable, which pressures enterprise infrastructure. And the open-source ecosystem—which provides the foundation everyone builds on—is more fragile than it looks. The Qwen situation is a reminder that the best research teams aren't factories. They're people. And people walk.