Hangzhou's Liwensuo opens Lévin Harness, an agent workspace for protein design

In mid-September the Hangzhou-based AI protein design company Liwensuo released a product called Lévin™ Harness and opened it to the research community, starting with support for Apple-silicon Macs. It is not a single model but an agent-centred protein design application: researchers describe a goal in natural language and the agent interprets the request, plans the steps and calls the right tools to run calculations and analyse results.

The problem it targets is the long-neglected last mile of protein design. To design a binder, a researcher must search the literature, inspect structures, download weights, configure environments and push jobs onto a server; after candidates come back, the work continues in other tools, viewing structures, filtering hits and judging which molecules deserve bench time. Models keep getting stronger, yet the wiring around them still has to be assembled by hand.

That reading matches the direction of the field this year. The article points to OpenAI's April life-sciences release GPT-Rosalind, Anthropic's Claude Science in June and Anthropic's August work on agents that take part in protein design and hand results to external experiments. The shared shift is that AI is no longer confined to the computation inside a model; it is joining the research work around it, reading results, revising plans and pushing a project forward.

In product terms, Lévin Harness places data, models, plugin tools, compute and workflows inside one agent workspace. General-purpose LLMs handle intent and planning, with the article naming DeepSeek, Kimi, GLM, Qwen and Xiaomi's MiMo, while specialised scientific models and analysis tools handle structure prediction and sequence design, among them AlphaFold, ProteinMPNN, the all-atom generator Pallatom and RFantibody for antibody design.

To let the agent see a molecule, the workspace embeds 3D molecular visualisation, so structure information on screen and the researcher's actions feed directly into the agent's context. A vague here in a conversation resolves to a specific spatial position, giving both sides a shared basis for checking residues and local interactions. A plugin system downloads, installs and invokes tools automatically; compute can run on a local workstation or be submitted to remote GPU servers for batch screening in the background; and a validated method can be saved as a Workflow for reuse and revision.

Liwensuo was founded in September 2021 and takes its name from the Levinthal paradox. The team first built the Lésign protein design platform and then pushed into all-atom protein generation with Pallatom, published in 2024 and selected for ICML 2025 Spotlight; NVIDIA's website has featured Pallatom as a representative protein foundation model alongside AlphaFold 3, ESM-3 and Proteína, and Pallatom-Ligand, aimed at small-molecule ligand-binding protein design, was accepted to ICLR 2026. The company built wet-lab teams alongside its algorithms from the start, and its Hangzhou industrialisation platform exceeds 3,000 square metres, including a 1,000-square-metre class-10,000 clean area.

Moving from an internal tool to a community product raises the question of data boundaries. Lévin Harness says a user's projects, conversations, settings and API keys stay on the user's own machine, conversations go to the model provider the user configures, backup is the user's decision, and Liwensuo will not train its own models on private conversations or files. Institutions can therefore adopt external intelligence while keeping control of their designs and research records.

The more interesting signal is how AI for Science is dividing labour: general models interpret goals, scientific models compute, labs verify, and the calls, combinations, judgements and feedback between those steps become the new bottleneck. Liwensuo says it wants Harness to keep connecting synthesis and experiments, edging toward what it calls a 24-hour intelligent R&D factory. For a company that grew out of models, that is both an extension of its product line and a bet that its internal know-how can be opened to far more laboratories.