Overview
Enterprise agents are attracting large investments and new control products, but evaluation remains sensitive to the harness around them. Microsoft is combining models for research, Okta is framing agents as identities, and a stricter WebVoyager protocol reports a substantially lower success rate than a vendor's earlier figure.
Developments
1. Sycamore raises $65 million for an enterprise AI agent platform
Enterprise AI agent startup Sycamore announced a $65 million seed round led by Coatue and Lightspeed. Founder Sri Viswanath, a former Coatue investor and Atlassian CTO, says the company intends to build and orchestrate complete enterprise solutions rather than add a single agent to an existing workflow.
The financing establishes investor appetite, not product-market fit. Sycamore declined to name the large customers it says it has, and it enters a market that includes startups, model vendors, and every major cloud. Any procurement case therefore depends on deployed references, explicit security boundaries, and practical exit options.
Sources: Sycamore's funding announcement · TechCrunch on Sycamore's seed round
2. Okta frames AI agents as a distinct identity class
Okta CEO Todd McKinnon told The Verge that enterprises need an inventory of agents plus roles, permissions, connections, monitoring, and a way to revoke access. He described agent identity as distinct from both a person and a conventional service account and called it a potentially large new security category.
Okta is defining a category it hopes to sell, and no market evidence yet identifies a winning identity model. The vendor-neutral architecture assigns each agent a unique, short-lived identity, binds permissions to a task, and keeps ownership, activity, and revocation visible without borrowing a human credential.
Sources: The Verge interview with Okta CEO Todd McKinnon · Okta on AI agent identity
3. Microsoft Researcher adds critique and model comparison
Microsoft added two early-access modes to Researcher in Microsoft 365 Copilot's Frontier program. Critique has Claude refine a GPT-produced response, while Model Council presents side-by-side outputs from Anthropic and OpenAI models and reports where they agree or disagree.
Multiple models can expose disagreement and catch some errors, but correlated training data and plausible shared mistakes limit the benefit. A sound evaluation checks final answers against sources, measures whether critique offsets its extra latency and cost, and retains model-specific logs for audit.
Sources: Microsoft's Copilot Cowork announcement · Engadget on Microsoft's multi-model Researcher
4. A stricter WebVoyager protocol lowers Operator's measured success
A March 30 preprint audits ambiguities in WebVoyager and introduces clearer rules for task instantiation, failure handling, annotation, and reporting. Its revised protocol reached 95.9 percent inter-annotator agreement and measured OpenAI Operator at 68.6 percent overall success, versus the 87 percent previously reported by OpenAI.
The gap cannot be assigned solely to Operator because benchmark wording, live websites, timing, and scoring all changed the measurement context. That is precisely the finding: a web-agent percentage is inseparable from task versions, retry policy, failure treatment, and annotator guidance, all of which belong beside a published score.