EvoMaster: 100 Lines of Code to a Scientific Agent
SJTU's SAI-Agents group just pushed EvoMaster's preprint and the v0.1.1 tag. The claim is that you can build a working scientific agent in about a hundred lines of code. The framework handles tool calling, skill composition, and memory, supports MCP out of the box, and plugs into an Anthropic-style Skills system. Apache 2.0. 118 stars and rising. It is the engine behind an actual family of deployed agents called SciMaster, including ML-Master, X-Master, and Browse-Master.
Why this matters. Most agent frameworks that claim to be minimal still make you wire up a router, a memory backend, a tool registry, retries, and a prompt layer before anything runs. EvoMaster collapses that stack because it treats skills and tools as the same primitive, and lets MagiClaw, a companion layer, generate new sub-agents on the fly without you writing orchestration code. The difference between a generic agent framework and a scientific-agent framework is that scientific agents need to read papers, call weird analysis tools, remember intermediate results, and then call more tools. EvoMaster is tuned for that shape.
The more interesting pattern. Chinese academic labs have figured out that to compete with industry agent frameworks, they need to ship real working systems, not toy demos. Tsinghua had ChatDev and AutoAgent. Shanghai AI Lab has InternLM agents. Now SJTU is pushing SciMaster plus EvoMaster plus an active roadmap including PhysMaster and EmboMaster. The production bar inside Chinese labs on agent tooling is quietly higher than most of what comes out of Western academia.
For anyone building a domain agent β legal, medical, scientific, finance β EvoMaster is worth thirty minutes of your evening. The 100-lines claim is real, and the preprint shows how the same primitive scales to a multi-agent scientific discovery loop.
Repo https://github.com/sjtu-sai-agents/EvoMaster
← Back to all articles
Why this matters. Most agent frameworks that claim to be minimal still make you wire up a router, a memory backend, a tool registry, retries, and a prompt layer before anything runs. EvoMaster collapses that stack because it treats skills and tools as the same primitive, and lets MagiClaw, a companion layer, generate new sub-agents on the fly without you writing orchestration code. The difference between a generic agent framework and a scientific-agent framework is that scientific agents need to read papers, call weird analysis tools, remember intermediate results, and then call more tools. EvoMaster is tuned for that shape.
The more interesting pattern. Chinese academic labs have figured out that to compete with industry agent frameworks, they need to ship real working systems, not toy demos. Tsinghua had ChatDev and AutoAgent. Shanghai AI Lab has InternLM agents. Now SJTU is pushing SciMaster plus EvoMaster plus an active roadmap including PhysMaster and EmboMaster. The production bar inside Chinese labs on agent tooling is quietly higher than most of what comes out of Western academia.
For anyone building a domain agent β legal, medical, scientific, finance β EvoMaster is worth thirty minutes of your evening. The 100-lines claim is real, and the preprint shows how the same primitive scales to a multi-agent scientific discovery loop.
Repo https://github.com/sjtu-sai-agents/EvoMaster
Comments