Benchling 工作坊实录：用 AI 智能体在 Opentrons Flex 上写协议、跑实验并分析数据

Hi Everyone,

At the summit's Build Day in June, the Benchling team brought their AI agent and a real Opentrons Flex into one room and handed both to a room of builders. The goal: describe an experiment in plain language, have the agent write the protocol, simulate it, run it on the robot, then analyze the data with the same agent.

The workshop edit is above. Below is the build guide the team wrote for the session, combined with what they said in the room, so you can follow along.

Why Benchling built this

Nari Kang, Head of Automation & Analytics at Benchling, opened with where the company is going. Benchling has spent the last decade helping scientists capture experiments and store their data, and now works with over 200,000 scientists. The next step is what they call AI scientists: tooling that works alongside scientists across the design, build, test and run loop.

Opentrons is one of Benchling's close partners, so scientists who already use Benchling can reach Opentrons protocols directly. The workshop zoomed in on the build-and-test part of that loop.

The setup

No pre-work. Everything ran inside Benchling AI:

Builders got an account on a workshop tenant at the start of the session.
Click the sparkle (AI) icon in the top left and ask it to list the Opentrons connections. You'll see the robot plus four simulators, one per row of tables, so nobody overloads the robot.
Optional: run your own Opentrons simulator locally (the Opentrons app, the opentrons repo, Docker and the Benchling Linux gateway) to watch your simulation run.

One thing the agent can't see on its own is what's on the deck. It knows which pipette arms are attached, but not which plates and tips are where, since that changes every protocol. So the guide gives you a layout to paste in:

50ul tips in C1, 1000ul tips in D1, 12 column reservoir in D2 with red, yellow, green, and blue color water in 3 slots each from left to right (RRRYYYGGGBBB) and a 96 well plate in D3

Design and execute

Nathan Stender walked through the first half. Opentrons protocols are just Python scripts, and Benchling's agent has a skill that writes them, sends them to the robot, simulates them to make sure they'll work, and runs them. It also monitors the run and reports back: running, finished, or an error it may need to fix.

His test right before the workshop: he uploaded a photo of the reservoir, told the agent what was loaded in D2, and asked it to write a protocol to transfer 20 microliters of liquid from the reservoir to the plate using the right mount tip. It wrote the protocol, uploaded it to the robot, ran the analysis and confirmed there were no errors.

The flow for builders:

Chat with the agent about what you want to do and what it can handle.
Push the protocol to your simulator. It checks the protocol for errors.
Once the analysis passes, raise your hand and run it on the real Flex, one person at a time.

You can treat the plate as a 96-well, 8 x 12 pixel grid and draw with colored water, or do something more scientifically realistic. At Benchling's Innovation Day the team built a protocol that let people pick a color and had the agent work out how to mix it.

Analyze the data

Janet Matsen took the second half. Her line: almost all of science is tabular data. Benchling has moved years of machine-learning pipeline work into the agent. You chat with it, it writes code that runs in a remote sandbox, and the results come back, with skills from working with many customers loaded in.

The workshop dataset was simulated plate data: compounds tested for efficacy across a 96-well plate, with replicates and positive and negative controls. As she put it, when you dump it in, you don't have to know anything about it. It's built for common open Python libraries today, not arbitrary bioinformatics pipelines.

In the room

Builders spent the hour writing protocols (color patterns, dilution series, a smiley face) and then crowding around the Flex to watch them run. The last minute of the video is that part.

Workshop team: Nathan Stender (Software Engineer), Janet Matsen (Software Engineer) and Nari Kang (Head of Automation & Analytics), Benchling.

Recorded at AI Science Summit 2026 Build Day in San Francisco. Presented by Worldwide Studios.

Stay tuned,

Michael

Ps. You're getting this note as you've joined a recent community event. If these updates are no longer for you, no worries, you can unsubscribe below.

注意转载授权 · 阅读原文