Today, we're introducing the Tamarind Agent Builder.
Every R&D organization has its own approach to molecular design, optimized for their targets, modalities and wet lab capabilities. Teams develop standards on which designs graduate for testing, how to assess designs, and which models perform best.
Agent builder lets you make your expertise into reusable, standardized agents.
Tamarind Bio brings a diverse set of scientific tools for your agents, including 400+ molecular design models, a best-in-class router to pick the best model for your task, epitope selection, literature review, and autonomous de novo design. Of course, your own skills, connectors, APIs, and tools are easy to onboard and standardize as well.
Tamarind handles the compute and orchestration, your team develops and refines the scientific knowledge. Build an agent once, then let your entire organization deploy it immediately.
Try out the Tamarind Agent Builder: app.tamarind.bio/agent
Define your agent in plain language
Build an agent by describing how it should work. Write its instructions in plain language, enable the capabilities it needs, including code execution, web search, pipelines, the molecule database, or fine-tuning, and choose its subagents, skills, MCP connectors, and effort level.
Capture your team’s expertise in skills
Skills capture your team’s expertise. Each is a Markdown file describing when it applies and what to do, with supporting scripts where needed. Our molecular dynamics skill, for example, instructs the agent to estimate trajectory size before submitting a job, choose the figure that answers the scientific question, and explain that a 2 ns simulation provides evidence of local stability, not binding or unbinding. Your skills might define developability thresholds, specify a numbering scheme, or encode the filters a design must pass before reaching the bench.
Choose models using relevant benchmark evidence
The model router uses benchmark evidence relevant to your inputs. Supply your chains and their roles, such as heavy, light, antigen, binder, or target, and it searches for your target or a related complex before ranking models. For trastuzumab’s Fv bound to HER2, it ranked ESMFold2 first for its performance on a related antibody–antigen target and low compute cost. It recommended OpenDDE as an independent prediction based on its overall antibody–antigen interface performance, followed by IntelliFold 2 as another alternative. It also made the limitation explicit: the exact complex had not been benchmarked, and the related example matched only on the antibody side. The benchmark therefore provided no direct evidence for the predicted antigen contacts. That uncertainty matters when deciding whether to build a design round around a structure.
Delegate research, analysis, and independent review
Agents can delegate specialized work to four subagents. A researcher investigates a focused question about a target, mechanism, or method using literature and public databases, with citations. A target analyst examines available structures to assess a known epitope or identify a binding site. A results analyst reads completed jobs and batches, calculates metrics, and assembles results tables. A blind reviewer independently evaluates raw outputs against the plan’s criteria before designs are finalized, without seeing the reasoning that produced them.
Bring your own tools
Your own methods can join the same workflow. Package a custom tool in a container, specify its GPU, memory, and runtime requirements, and deploy it through the UI, CLI, SDK, API, or a GitHub repository. Once published, it appears in the catalog with the same job schema as the other tools, so agents and pipelines can call it just as they call AlphaFold. Builds are versioned, and you can roll back by republishing an earlier build.
Version and share agents across your organization
Agents and skills are also versioned and published to your organization, with a record of who published each version and when. Published agents pin their skill versions, keeping their behavior consistent as your team develops new methods. The platform is exposed over MCP, so the agents you build are available through the Tamarind app, the API, and MCP clients including Claude.