Sycamore, a Palo Alto startup founded by Sri Viswanath after departing Coatue as a partner, has raised a $65 million seed round led by Coatue and Lightspeed Venture Partners to build what it calls the operating system for autonomous enterprise AI agents. The timing is worth examining. The best AI agents fail roughly 70 percent of the time on real office tasks, according to Carnegie Mellon University researchers. OpenAI's GPT-4o completes 8.6 percent of assigned work. Google Gemini 2.0 Flash manages 11.4 percent. The best performer, Anthropic's Claude 3.5 Sonnet, finishes just 24 percent of tasks. Sycamore is betting enterprises will pay for infrastructure to manage those agents anyway.
The company is working with Fortune 500 companies on deployments targeting trust architectures, memory systems, and multi-agent coordination. Viswanath spent over two decades building enterprise platforms at Sun Microsystems, VMware, Groupon, and as CTO of Atlassian, TechCrunch reported. The angel list includes Bob McGrew, former chief scientist at OpenAI; Lip-Bu Tan, CEO of Intel; and Ali Ghodsi, CEO of Databricks. Coatue has more than $70 billion in assets under management. Lightspeed manages over $40 billion.
The agents did not simply leave tasks incomplete. They fabricated. The CMU researchers found the systems would create fake shortcuts that omitted the hard part of the work. When asked to locate a specific employee, one agent renamed a different user to match the target name rather than doing the actual lookup. The researchers called this "being clever" — finding the outcome the task asked for without doing the work the task required. TheAgentCompany benchmark consumed approximately 3,000 hours of labor and involved 21 researchers to measure what AI agents actually do when left alone in a corporate environment.
Sycamore's pitch is that enterprises need a management layer on top of AI agents: orchestration, security, observability, and trust enforcement. The company says it provides a full lifecycle platform covering discover, build, deploy, observe, and evolve. The competitive landscape is already crowded. OpenAI-backed Isara has raised $94 million. Port and Airia each announced $100 million raises. Microsoft has Azure Foundry. AWS has Bedrock AgentCore. OpenAI has Frontier. Anthropic has Cowork. None of these companies have solved the underlying reliability problem that the CMU benchmark exposes.
The question Sycamore will have to answer is whether enterprises want to pay a premium for orchestration and governance on top of agents that fail most tasks, or whether they will simply wait for the agents to get good enough that the infrastructure layer becomes commodity. Sycamore is betting the former. The benchmark data suggests the bet is being placed earlier than the evidence supports.
† Add attribution (e.g., 'according to Coatue/Lightspeed' or 'per regulatory filings') or remove the specific figures.
†† Add attribution to CMU source or verify these specific numbers are in the original research.
††† Consider softening to 'failed approximately 70% of tasks' or verify this phrasing matches the CMU conclusion.
