The Sakana Marlin agent takes 8 hours to generate reports
Marlin, a self-operating AI agent that generates 60–100 page researched reports with references and slides, has been released by Sakana AI.
Sakana AI’s latest product, Marlin, is a self-operating enterprise research AI agent which autonomously performs tasks for up to around eight hours, generating 60 to 100 page long researched reports with references and slides. It’s described as a “virtual CSO”, as it runs through hundreds to thousands of model queries while planning hypotheses, searching information and validating results, all within a single session, saving weeks of strategic planning time.
Marlin was made available for purchase starting in June 2026, following a closed beta period from April. During this time, approximately 300 people were able to test out the product. It operates on a credit system, with each use costing 100 credits at a pay-as-you-go price of ¥98 per credit, or ¥150,000 per month for the Pro plan (2,000 credits), ¥400,000 per month for the Team plan (6,000 credits), and custom prices for enterprises.
Why it can run for hours, not seconds
This product uses a search technique known as AB-MCTS (Adaptive Branching Monte Carlo Tree Search). Rather than giving a single response, it approaches thinking as a tree structure, choosing between expanding the search space by creating new candidates, or deepening the search by improving existing ones. There’s also a multi-model version which uses various Large Language Models (LLMs) for different parts of the process, such as o4-mini, Gemini 2.5 Pro and DeepSeek-R1. According to Sakana, this allows them to solve around 27.5% of the difficult ARC-AGI-2 benchmark, compared to 23% when using o4-mini by itself. The underlying search algorithm is open sourced under the Apache 2.0 license as TreeQuest, so you can implement the concept without needing to use Marlin itself.
Why this matters to you
Unlike most AI products that give quick responses and require active guidance throughout the process, Marlin represents a new wave of AI assistants that you can set a task for and leave alone for several hours. This shifts the focus from individual interactions to finished products, and from “can it write?” to “can I trust a report written by AI without seeing how it was done?”. Unfortunately, right now the answer is no, we can’t fully trust it. We’ve already seen a large company retract an AI-generated report due to false information.
In my experience with long-term agents, I found that the eight hour runtime is true and can help create a rough draft, but to get the most out of it you need to think of the result as something to build upon. The 60–80 referenced sources are where the real value lies, because those are the things you can check. Without doing so, you’re just automating the creation of confidently wrong, well formatted text.
Use a long-running agent for the legwork, not the verdict. Before anyone acts on a Marlin-style report, click through the cited sources for every claim you'd stake a decision on, and never forward a document you haven't read end to end just because it looks finished.
Source: www.marktechpost.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles