Forget the Model: the Agent Harness Is the Real Product Now

Google turned Antigravity into a product line while Alibaba and xAI tuned models for 1,000-call agent runs. The orchestration layer is the new battleground — and your harness choice now matters more than your model.

The race among AI labs this week wasn't about a model topping a benchmark — it was about the scaffolding around the model. Google spent the week turning Antigravity from a feature into a product line, and the subtext is hard to miss: the orchestration layer that drives an agent through dozens of tool calls is becoming the thing labs actually sell.

The harness becomes the product

Google DeepMind said it is expanding the Antigravity ecosystem with 3 new entry points — a "mission control" 2.0 where agents work a project simultaneously, a terminal CLI, and an SDK — alongside a $2M Build with Gemini XPRIZE hackathon (Google DeepMind, May 19, 2026). Yet even seasoned developers aren't sure what they're buying. Simon Willison asked publicly whether "Antigravity" is a generic label for Google's agent harnesses or a single closed-source binary you download. That ambiguity is the story: when the harness is the product, naming and lock-in matter as much as model quality.

The competitive logic was named most cleanly by swyx, who tied the rise of "Agent Labs" to a direct correlation between model performance and agent-product revenue, with a discontinuity he places in Q4 2025. The business that improves as the model improves, in other words, is the harness wrapped around it.

Everyone is now tuning models for that harness. Alibaba introduced Qwen3.7-Max for long-running agent work; in an internal test the company claims the model made more than 1,000 tool calls to optimize a compute kernel, improving inference speed roughly 10x over the prior version (TechNode, May 2026) — a vendor claim, not an independent benchmark. xAI opened Grok and X Premium access inside the openclaw harness, and its Grok 4.1 Fast model (already marked deprecated in Oracle's docs as of May 2026) advertised a 2 million-token context window and roughly 3x fewer hallucinations, both aimed at multi-step agent tasks.

What builders and buyers should do

For builders, the takeaway is uncomfortable: choosing a model is now also choosing a harness, and the two arrive bundled. Portability — swapping Gemini for Grok or Qwen without rewriting your agent loop — is the thing to protect. For buyers, the benchmark that matters is no longer a single answer's quality but how far an agent runs unattended; treat "1,000 tool calls" as a claim to reproduce with your own eval, not a spec to trust.

The bottom line: Pick your agent harness as deliberately as your model this quarter — a model-agnostic layer you can re-point at the next benchmark leader beats a vendor-bundled stack you'll have to re-platform.

Sources