ScienceClaw Explained: MIT's AI Agents With No Boss
MIT's LAMM lab built a framework where AI agents run science with no central planner. Here's how the needs board works, what the four case studies show, and where it stops short.
- ScienceClaw + Infinite is an open-source framework from MIT's Laboratory for Atomistic and Molecular Mechanics (LAMM), posted to arXiv on March 15, 2026.
- There is no planner model. Agents broadcast unmet needs to a shared index, and peers claim them through a pressure score in a component called the ArtifactReactor.
- Every skill call becomes an immutable artifact with parent lineage, so each finding traces back to raw tool output.
- Four case studies ran with 8 to 13 agents each. Coordination only works on a single machine, and the paper has no head-to-head against a centralized planner.
ScienceClaw is an open-source framework from MIT's LAMM lab where AI agents run scientific investigations with no central planner. Agents post unmet needs to a shared index, peers claim them by pressure score, and every tool call is logged as an immutable artifact with its lineage, then published to a platform called Infinite.
Every agent stack I’ve shipped has a boss. One planner model reads the request, splits the job, and hands out tickets. If that planner hallucinates the plan, every agent downstream executes a bad plan with perfect discipline.
MIT’s Laboratory for Atomistic and Molecular Mechanics (LAMM) published a framework that fires the boss. ScienceClaw + Infinite hit arXiv on March 15, 2026, and it’s the cleanest look I’ve seen at the AI systems race with the orchestrator pulled out.
I haven’t run it. This is a read of the paper and the open-source repo, the receipts they report and the gaps they don’t cover.
What is ScienceClaw?
ScienceClaw is a framework where independent AI agents run scientific investigations with no central coordinator. The paper lists seven authors, all at LAMM, with Markus J. Buehler as corresponding author. The code is open source on GitHub under Apache 2.0.
Three pieces do the work. First, a skill registry the repo pegs at 300+ tools: PubMed, BLAST, UniProt, AlphaFold, PDB, PubChem, RDKit, Materials Project and more.
Second, an artifact layer. Every skill call produces an immutable record with a UUID, a content hash and its parent lineage. Third, Infinite, a platform where agents publish findings as structured posts that humans and other agents can vote on, cite and build from.
Each agent gets a profile, basically a scientific personality plus a preferred tool list. A biology agent and a chemistry agent on the same topic reach for different tools. That’s the point. New to the term? Start with what an AI agent actually is.
How plannerless coordination works
Here’s the move. When an agent needs something it can’t produce itself, say a structure prediction, it doesn’t ask a manager. It broadcasts the unmet need to a shared global index.
Peers run a component called the ArtifactReactor. It scans that index and ranks open needs by “pressure.” The paper builds pressure from four inputs: novelty, centrality, depth in the lineage graph, and age, so old needs drift up and nothing starves.
Think pickup ball at the park. Nobody’s coaching. You see the open spot, you call next, you play the position your game fits.
When two or more compatible results land for the same need, the reactor merges them into a multi-parent synthesis. A mutation layer prunes branches that conflict or repeat. Everything lives in a directed acyclic graph, so any finding traces back to the raw tool call behind it.
The rhythm comes from a heartbeat daemon. Per the README, it wakes each agent every 6 hours for a full cycle.
What the four case studies show
The paper runs four autonomous investigations. Peptide design for SSTR2, a receptor overexpressed in most neuroendocrine tumors. Lightweight impact-resistant ceramics. A resonance study spanning biology, engineered acoustics and music. And a formal analogy between city street-network growth and grain-boundary evolution.
Table 1 is where the receipts live. Protein design: 10 agents, 23 tools, 177 artifacts. Resonance: 13 agents, 12 tools, 159 artifacts. Materials: 8 agents, 10 tools, 73 artifacts. Formal analogy: 9 agents, 23 tools, 52 artifacts.
The ceramics run is the cleanest result. Agents screened 212 light-element ceramic phases. 14 cleared both bars, density under 5 g/cm³ and bulk modulus over 200 GPa, and 7 were thermodynamically stable. Boron carbide (B4C) and boron suboxide (B6O) came out as the strongest lightweight picks.
The peptide run is the one I respect most, because it graded its own homework. The optimized candidate, MGLKNFFLKTFTSC, beat the seed on predicted fitness. Then the analysis flagged that it weighs about 1639 Da, way heavier than octreotide at roughly 1019 Da.
A system that catches its own problem beats one that ships a pretty number and walks off.
Where it falls short
Now the part the hype posts skip.
It’s decentralized on one box. The README says it straight: emergent coordination currently works within a single machine, because every agent reads the same local index file. Cross-machine coordination is listed as planned. So “collective” is a design pattern right now, not a network of strangers’ agents trading work.
No head-to-head. The paper doesn’t benchmark against a centralized planner. No cost figures. No run times. Anybody telling you a swarm beats one big model on price or speed because of ScienceClaw is making that up.
The counts don’t always agree. The skill registry shows up as “over 300,” “270+” and “200+” in different sections. The resonance study is ten agents in the text and 13 in Table 1. Small stuff. Still worth knowing before you quote it.
What builders should steal
You don’t need to design peptides to use this. The pattern travels.
One: a needs board. Let agents post what they’re missing instead of routing everything through one planner prompt. Two: lineage on everything. Hash every tool output and record its parents, so an audit is a graph walk, not an archaeology dig. Three: pruning. Collectives create duplicate work, so build the cleanup in from day one.
It’s model-agnostic too. The repo runs OpenAI by default, plus Anthropic, Gemini and Hugging Face backends, and the paper’s main deployment used Claude Opus 4.6/4.5. If you’re already stacking tools from the best AI agents of 2026, this is a blueprint for wiring them together without one model playing traffic cop.
ScienceClaw doesn’t prove the boss model is dead. It proves you can run real investigations without one and keep every receipt.
Stop hiring a manager for your agents. Give them a job board and a paper trail.
#TheAIMogul
Bottom lineScienceClaw doesn't prove the boss model is dead. It proves a crew of agents can run real investigations without one and keep every receipt. Steal the needs board and the lineage log, and don't believe anyone selling it as cheaper or faster, because the paper never measures that.