Sakana AI's 'Marlin' Agent Can Research Autonomously for 8 Hours — Compressing Weeks of Strategy Work Into a Single Session

Sakana AI, the Tokyo-based research lab founded by former Google researchers David Ha and Llion Jones, has unveiled Marlin — branded publicly as "Sakana Ultra Deep Research" — an agentic AI system designed to automate extended business research and strategy work.

The core capability differentiator is runtime: Marlin operates autonomously for up to eight hours on a single task, substantially longer than typical AI assistant sessions that time out or lose coherence after minutes of operation. During that window, the system independently searches, synthesizes, cross-references, and structures information into finished strategic deliverables.

What Marlin Actually Does

According to Sakana AI's launch materials, Marlin is specifically designed for business intelligence tasks that currently require teams of human analysts working over days or weeks. The target use cases include competitive landscape analysis, market entry assessments, technology due diligence, and regulatory environment mapping.

The system accepts a high-level strategic question from a user — for example, "Assess the competitive dynamics and market opportunity for AI-powered ERP systems in Southeast Asia over the next three years" — and returns a structured, sourced analysis without requiring further human intervention during the research process.

Marlin is currently in beta testing with a limited group of business customers. Pricing and general availability details have not yet been disclosed.

The Agentic Research Race

Marlin enters a rapidly crowding market. OpenAI's Deep Research, Google's Gemini Deep Research, and Perplexity's research features all target the extended autonomous research use case. The differentiators in this category are increasingly around depth of synthesis, citation quality, and the ability to handle ambiguous or multi-step analytical tasks rather than simple information retrieval.

Sakana AI's distinctive positioning is its research heritage — the lab has published influential work on evolutionary model merging and self-improving AI systems. Marlin appears to leverage multi-agent orchestration techniques developed in that research context, with individual sub-agents responsible for different dimensions of a research task running in parallel before a synthesis layer assembles the final output.

For knowledge-intensive industries — consulting, investment banking, corporate strategy, policy research — the prospect of compressing weeks of analyst time into hours represents a genuine workflow disruption. Whether Marlin delivers on that promise at the quality bar those industries require remains to be validated.