Benchling Science Foundries Takes AI Agents Into the Lab

The next great workplace for AI agents might be a wet lab. On October 15, Benchling unveiled its new product, Benchling Science Foundries, a new orchestration environment that puts teams of AI agents in charge of entire research and development workflows — from designing molecules to ordering materials from contract research organizations to running lab equipment — while scientists decide where the agents stop and human approval begins. According to the company's launch announcement, the product represents a sharp escalation of Benchling's agentic strategy: agents have become the fastest-growing capability in the company's history, and Science Foundries now extend them from discrete lab tasks to whole scientific processes.

The timing matters for anyone watching the agent economy. Across enterprise software this year, the question has shifted from what an individual AI agent can do to how teams of agents coordinate toward a shared goal inside real organizational constraints. According to reporting by R&D World, Benchling introduced Science Foundries at its Benchtalk conference in Boston, pitching a system where five agents — coordinator, design, cloning and sourcing, lab planning and data, and analysis and learning — run on daily schedules or event triggers, all working from the same scientific data and context. A timeline view shows their activity around the clock, including work done near midnight while the scientists sleep.

From discrete tasks to end-to-end workflows

Before this launch, Benchling's agents handled individual lab tasks inside its scientific data platform, which is used by more than 1,300 biotech companies. Science Foundries change the unit of work from the task to the process. According to the company's launch announcement, an R&D team picks a goal — produce a set of proteins, optimize a promising lead, improve a manufacturing process — then decides which steps agents will run and where scientists need to approve. From there, the Foundry runs on its own: agents execute and monitor work, coordinate handoffs between steps, and surface only key decisions to scientists.

The flagship example is a pharmaceutical team designing, producing, and testing antibodies at high throughput. The team already had the pieces: design models, contract research organizations to synthesize constructs, and lab automation for production and assays. The missing piece was connection — scientists were moving data by hand from one step to the next. As reported by R&D World, the team's Foundry now splits the work across four specialist agents. A design agent generates and evaluates sequences using scientific models. A cloning agent drafts assembly strategies and places orders with a CRO. A lab agent sends instructions to workcells, captures the data, and runs quality control. An analysis agent interprets results and updates the training datasets so the next round gets smarter.

Benchling frames the benefits in two ways. The first is speed between steps: an assay that finishes late at night gets analyzed before the scientist arrives in the morning. The second is breadth: agents can run several design strategies against the same goal in parallel, where a team previously tested one approach at a time. Company co-founder and president Ashu Singhal argued at the launch that running lab automation, CROs, and generative models in parallel toward one scientific goal is a new kind of coordination challenge — one that agent teams are designed to solve.

The approval points are the product

Look past the agent choreography and the most interesting design decision is the governance model. Science Foundries ask scientists to declare, up front, which steps agents run and which require human sign-off. That makes the human checkpoint a first-class part of the workflow rather than an emergency brake bolted on afterward.

This trust architecture is spreading across the enterprise agent stack. ServiceNow's AI Passport gives every enterprise AI agent a signed identity so its actions can be traced (identity first), and RapidCanvas's data science agents automate the modeling pipeline for business teams (agentic data science). Benchling's version of the pattern is domain-specific: in a regulated lab, an agent that can order from a CRO or instruct a workcell needs its autonomy bounded by scientist approval points, or nobody downstream can trust the experiment.

The pattern is converging because the failure modes are converging. An agent that acts on incomplete context in a lab can waste expensive reagents or misdirect a weeks-long synthesis. In every vertical, the answer is becoming the same: bounded autonomy with explicit checkpoints, auditable actions, and humans who remain accountable for the outcome.

Biotech is the stress test for agent teams

There is reason to take the lab use case seriously — and reason to stay skeptical. On the serious side, biotech R&D is exactly the kind of work that breaks naive agents: long-running processes, physical-world consequences, multiple external organizations in the loop, and data that lives in instruments, CRO reports, and scientists' heads. A system that can orchestrate all of that toward a shared goal is doing the hardest version of agentic coordination.

On the skeptical side, the launch arrives without measured customer results, as noted by analysts covering the announcement. Whether Foundries work in practice will depend on model reliability, integration with lab systems, and access to experimental history — the last of which is where Benchling holds a structural advantage, since its customers have accumulated years of structured scientific data on its platform. According to coverage by BiopharmaTrend, Benchling envisions labs where scientists oversee multiple AI agents working with models, robots, and external providers, reviewing results and deciding which experiments to pursue next rather than manually coordinating each step.

Benchling has been building toward this for 18 months of AI investment, and the company says agents are the fastest-growing capability it has ever released, with scientists already delegating real work daily. The Benchling Science Foundries bet is that the next order of growth comes from letting agents run whole workflows — not just isolated tasks — under scientist supervision. For the agent ecosystem, the lab is becoming a proving ground: if teams of AI agents can coordinate protein production across models, CROs, and robots with humans holding the approval keys, the same orchestration patterns will migrate to every other domain where work has real consequences.