The Agent Fight Just Moved Off the Model

Google, xAI, Alibaba, and indie devs all shipped agent scaffolding this week — not new models. The orchestration "harness" is becoming the real product, and the model is turning into swappable config.

The interesting fights in agentic AI this week weren't about models — they were about harnesses, the scaffolding that wraps a model and turns it into an agent that plans, calls tools, and spawns subagents. Google, xAI, Alibaba, and independent developers all shipped at that layer within days of each other, which suggests the model itself is becoming a swappable component while the value migrates to the orchestration around it.

The clearest tell is that nobody agrees on what to call this layer yet. When Google described Gemini Spark as running on Gemini 3.5 and using "the Antigravity harness," developer Simon Willison openly asked whether Antigravity is a generic term for Google's agent harnesses or a specific downloadable binary — the category is too new for settled vocabulary. Google DeepMind then showed Gemini 3.5 Flash inside Antigravity deploying multiple subagents to design a simulated city, the same subagent pattern Codex users are now scripting for deep research. xAI made its move by letting Grok and X Premium subscribers plug into OpenClaw, distributing its model through a third-party harness rather than mandating its own.

The capability frontier is moving in step. Alibaba's Qwen3.7-Max, announced this week (May 2026), is aimed explicitly at long-running agent workloads: in an internal test the company says the model made more than 1,000 tool calls to optimize a compute kernel, claiming a roughly 10x inference-speed gain. That is a vendor claim, not an independent benchmark — but it marks where "agentic" now points, toward hours of autonomous iteration rather than a single reply.

For builders, the practical shift is that you increasingly pick a harness before a model. Simon Willison's Datasette Agent alpha makes the point from the open-source side: it runs in the browser, swaps between models, and works with local options above 4B parameters like Qwen 3.6 — the harness is the product, the model is configuration.

For buyers, swyx's "Agent Labs" thesis — that the businesses worth backing improve automatically as models do, with a revenue discontinuity he dates to Q4 2025 — explains the rush to own the compounding layer. Alibaba opening its full Taobao catalog to Qwen agents shows the destination: agents that don't just answer, they transact.

The bottom line: Choose your agent harness deliberately this quarter — it is becoming the load-bearing decision, with the underlying model an increasingly swappable detail.

Sources