Artificial intelligence models are increasingly integrating what were once external tool functions, yet the infrastructure supporting these models—known as the harness layer—is not shrinking but expanding. Industry insights reveal that the traditional notion of a model as just a set of weights no longer holds. Instead, AI now operates within a broader ecosystem of tool invocation, hosted search, code execution, containers, and agent harnesses that enable complex workflows.
This harness layer has become essential for managing tasks such as context handling, sandboxing, permission controls, memory, checkpointing, and cost management. The trend shows that while models absorb certain scaffolding aspects, new levels of scaffolding emerge to address higher-level needs. This dynamic creates a continually shifting frontier rather than a diminishing support role for the harness.
Google’s Antigravity harness, developed from the Windsurf team, exemplifies the increasing importance of these systems. It serves as a connective platform uniting Search, the Gemini app, Cloud, and AI Studio, reflecting strategic investments in harness technology. These investments respond to the reality that robust long-running work, such as comprehensive coding, demands an integrated product framework beyond just model improvements.
The competitive landscape now favors teams that deliver integrated products combining models, tools, execution environments, and harnesses. Startups focusing on vertical domain expertise often outperform broader model labs, leveraging specialization as a key advantage. This shift has changed the unit of competition in AI development from models alone to complete AI solutions.
Looking ahead, the main challenges for AI performance will revolve around reliability and calibration over extended tasks, rather than incremental benchmark improvements. Current leaders in this area include labs like Anthropic and OpenAI, which have made strides in ensuring stable performance in long-running processes—an area where simply increasing model size yields diminishing returns.
Moreover, ownership of agentic products creates a data flywheel effect, strengthening training capabilities. Labs lacking such integrated products face growing barriers to entry as valuable assets become scarcer. These dynamics are compounded by policy factors such as U.S. export controls, which restrict the global reach of American AI labs. Simultaneously, Chinese open-weight models face fewer such limitations, potentially leading to a bifurcated market where American systems remain most capable but Chinese deployments dominate in scale.
Economic considerations also play a major role, with high-performing systems carrying significant operating costs, while more cost-effective open models offer slightly lower performance. How these cost-performance trade-offs evolve will further shape AI accessibility and market distribution over the next year.