Sakana's MASS: one model splits into subagents, grades itself, and gets measurably better

Good morning 👋 Sakana let one model split itself into a team, watch the team work, then decide which version of itself won. Two benchmarks said yes. Two said not yet.

In today's issue:

• 🔭 Sakana's MASS: one model splits into subagents, grades itself, and gets measurably better
• 🔬 Agents that branch instead of restart, and temperature 0 that still isn't reproducible
• 🛠️ Splitting one model into three drafts and letting it judge the winner
• 🔗 Tencent's $5B AI bond, FTX's old bitcoin mines turned data centers, and LegalOn's 65% Codex cut

🔭 THE ONE THING

🪞 Sakana's MASS: one model splits into subagents, grades itself, and gets measurably better

Sakana AI released MASS (Multi-Agent Self-Supervision): one shared model spins up virtual subagents, searches for better multi-agent workflows, then judges its own output to pick which ones to keep. Run on a 27B open-weight model, two rounds of that loop produced real score-per-token gains, 1.60x on MLR-Bench, 1.35x on DSBench, 1.28x on AstaBench, 1.20x on ScienceAgentBench, with paper and code both public.

Sakana's actual claim is narrower than "the agent got smarter." It's that the system has been bottlenecked on optimizer capacity, how well it searches for workflows, not on its capacity to judge them. Worth the asterisk: on Terminal-Bench 2.0 and SWE-bench Verified, the same self-improve loop dragged score-per-token to 0.94-0.96x baseline. Real gain, narrow one, confined to research-style agent tasks for now, not your coding agent.

🔬 RESEARCH HIGHLIGHTS

• Sandbox snapshots for coding agents Most agent harnesses still retry from scratch after a failure, throwing away whatever progress the agent made. A new ArXiv paper treats that as a straight waste: snapshot the sandbox mid-task, then let the agent branch from the last good state instead of redoing everything. Builder read: if you're running agentic coding loops in production, retry-from-zero is probably costing you more compute than it needs to. arxiv.org/abs/2610.13841

• Temperature 0 doesn't mean reproducible Semantic query engines ask an LLM to judge rows one at a time, and most benchmarks score them on a single pass. This paper re-ran the same workloads from SemBench, LRO-Bench, and UDA-Bench and found real answer drift run to run, even at temperature 0. Watch, don't act: if your eval pipeline only runs once, you may be grading noise instead of quality. arxiv.org/abs/2610.13823

🛠️ TRY THIS

Draft a doc three ways, then let the agents grade their own homework

MASS's trick, one model splitting into subagents and judging which one actually worked, scales down fine to your laptop this week. Genspark's genoffice gives Claude Code (or Codex, or Cursor) a CLI that produces real .docx and .xlsx files instead of markdown approximations, so the split-judge-pick loop ends in something you can actually send.

1. Run npx skills add genspark-ai/genoffice to drop the agent skill into Claude Code.

2. Give it one brief and have it spin up three subagents, each drafting a different outline of the same report as a separate .docx.

3. Run a fourth pass that scores all three against your own rubric (structure, clarity, does it answer the brief) and names a winner.

4. Take the winner's structure, graft in the single best paragraph from a runner-up, and export the final .docx with genoffice.

Prompt: Using genoffice, draft three outlines for [report topic] as separate .docx files, each from a different angle. Score all three against this rubric: structural logic, clarity, and whether it actually answers the brief. Name the winner and why, then merge in the strongest paragraph from a runner-up.

Worth a look

• T3MP3ST elder-plinius's new red-teaming harness runs the same split pattern on offense: spin up attacker agents, judge which exploits land, 6.5k stars already.
• AWS SageMaker HyperPod if running three subagents at once starts eating shared GPU, AWS published a reference setup for splitting one cluster across teams without anyone starving the others.

🔗 QUICK LINKS

• Tencent is pricing a roughly $5B bond, split between dollar tranches and offshore yuan notes, with the bulk earmarked for AI chip capex. It's the second mega-raise since a $4.7B bond in June. Bloomberg
• SWI Group raised €800M to turn former FTX-linked bitcoin mines (Genesis Digital's old sites in Texas, the Carolinas, and Sweden) into AI data centers. The founder told WSJ he has "zero interest" in mining bitcoin anymore. WSJ
• LegalOn cut its daily Codex bill 65% just by routing coding tasks across OpenAI's Astra, Sol, and Luna instead of defaulting to one model. OpenAI

That's the one worth sitting with today.

Pradeep Perugu