a16z Partner Fabrizio Serafini: Calls for Scalable Computer Use Agents
a16z partner Fabrizio Serafini posted on social media, pointing out that after decades of softwareing business processes, companies still rely on manual copy-pasting between multiple browser tabs for system connections, referring to this as "human APIs."
He explicitly requested startups to develop "truly effective" computer use agents, demanding that these agents be trustworthy and capable of executing real tasks at scale, directly operating within the company's existing software rather than forcing the creation of new interfaces.
Serafini joined a16z around May 2026, focusing on AI application investments, having previously worked as an analyst at Insight Partners and graduated from Harvard University with a degree in statistics and computer science. He previously co-authored a report showing that computer use agents improved from about 42% to the current best of approximately 85% on the OSWorld benchmark, surpassing human testers' level of about 72%.
a16z's speedrun accelerator is currently open for SR008 applications (priority window until November 1), offering up to $1 million in investment ($500,000 for 10% SAFE, plus up to $500,000 in follow-on investment), along with over $10 million in cloud and software credits, in a 12-week San Francisco-based program. This program has cumulatively invested hundreds of millions of dollars in several early-stage companies.
In his post, Serafini tagged a16z and speedrun, included a video explanation, and invited founders building such agents to reach out, emphasizing that the agents must be able to reliably complete end-to-end tasks in existing software environments.
Investors are directing capital towards computer use layers that can replace "human APIs"; event-driven applications are the gap after model capabilities cross the deployable threshold; funding is flowing to early-stage agent infrastructure and vertical workflow companies, benefiting teams that can deliver trustworthy scalable agents, while traditional system integrators and inefficient business processes relying on manual integration are under pressure.
Source: Public Information
ABAB AI Insight
Fabrizio Serafini graduated from Harvard with a degree in statistics and computer science and joined Insight Partners for growth-stage due diligence and strategy before joining a16z's AI application team in 2026. He has publicly analyzed the data changes of computer use agents from demonstration to production. a16z continues to invest in early AI and enterprise application companies through the speedrun accelerator, including multiple rounds of agent-related projects.
In terms of capital pathways, a16z is concentrating funds and network resources through public requests and accelerator terms on computer use agents that can perform real work on existing software (rather than new APIs). The motivation is that models have surpassed human levels on benchmarks, but companies still rely on manual bridging systems, presenting a scalable replacement opportunity; this is manifested in directly soliciting and providing up to $1 million in early capital and follow-on rights.
Comparable cases include early RPA vendors like UiPath in screen automation, as well as current computer use tools provided by Anthropic, OpenAI, and Google, but those are more model-layer oriented or require building environments. The industry is currently in a phase of expanding from model capability validation to deployable, trustworthy scalable agents, with speedrun directing resources towards early teams that can fill this gap.
Essentially, this is a technological replacement: computer use agents are replacing the role of "human APIs" in manually copying and pasting between multiple systems. The mechanism is that model visual and action capabilities have crossed the threshold for production usability, and companies are unwilling to rewrite APIs for every system, thus automating low-value manual bridging through screen-layer automation into scalable agent execution, shifting capital from human integration to agent infrastructure and vertical workflows.
ABAB News · Cognitive Laws
What models can do does not equal what enterprises can trust
Systems without APIs are the real battleground for agents
Accelerator requests equal the gap being priced by capital