Berkeley Lab’s TEM Agent: wrap the instrument, don’t train the model

Short answer: Researchers at Lawrence Berkeley National Laboratory (Berkeley Lab) built TEM Agent, which runs the TEAM 0.5 electron microscope, which Berkeley Lab calls one of the most advanced microscopes, from plain-English requests (Berkeley Lab). It uses an off-the-shelf model, bounded tools and a human in the loop, and the code is open source. The work is peer-reviewed. The authors report demonstrations, not benchmarks. We predict AI augmentation like this will become the default standard in the lab. The lesson for builders is the architecture: wrap the instrument's interface, don't train a model.

How does TEM Agent work?

  • Off-the-shelf model. TEM Agent uses Anthropic's Claude Sonnet 4.5 via Claude Code.
  • Four tool servers using the Model Context Protocol (MCP): the microscope, its 4D camera, the Crucible metadata system, and the Distiller pipeline at the National Energy Research Scientific Computing Center (NERSC) (code on GitHub; Crucible).
  • Peer-reviewed: "TEM Agent: enhancing transmission electron microscopy with modern AI tools," by Morgan K. Wall, Alexander J. Pattison, Edward S. Barnard, Stephanie M. Ribet and Peter Ercius of Berkeley Lab's Molecular Foundry, npj Computational Materials, June 10, 2026 (doi). The Oct 29, 2026 release publicized this earlier work.

Can an AI agent run an electron microscope safely?

The design aims for that. The model never writes code for the instrument. It can only call tools the team built, designed so that even misuse "would not damage the microscope." That's a design claim, not a tested guarantee.

What the authors don't claim

They're unusually direct about limits:

  • No success rates or time savings were measured. These are demonstrations.
  • It's non-deterministic. Repeated prompts gave different action sequences.
  • It depended on years of prior code and metadata.
  • "Traditional software ... would likely outperform any agent in the near term."
  • "There is no replacement for an experienced human operator."

That's why we frame this as augmentation, and as a prediction, not a finding.

Why it matters

Vendor openness may decide which labs automate. This approach works only where the instrument maker exposes a scripting interface and good metadata already exists. Off-the-shelf models plus MCP may be the cheapest route to agent-run instruments, but only for open instruments (our view).

The safety pattern might travel. Bounded tools are a different philosophy from Neuron Industries' approach in factories, where AI generates control code and then tests it in simulation. Whether bounded tools work in a factory or robot cell hasn't been shown.

What it means for you

  • Engineers in AI and ML: this is directly usable. Clone the code and study how the tools are bounded.
  • Founders building agents or integrations: the opening is in wrapping instruments, metadata and permissions, not in models.
  • Investors: it's a moat test. Value may sit with whoever owns the instrument interfaces and the metadata, not the model (our view).
  • Defense test and evaluation, and AI assurance: a bounded-tool design is likely easier to authorize than an agent that writes its own code.

Open questions

  • Has anyone measured TEM Agent's success rate or time savings?
  • Does the bounded-tool pattern hold up in factory cells?

Your turn: which of your instruments expose a scripting interface? If you've wrapped one for an agent, reply and tell us how it went. Follow LA Robotics Network for more on AI in the lab.

Previous
Previous

Neuron’s $10M seed: AI-written PLC code looks promising. Dark factories are further off.

Next
Next

Government is buying robot labs. Ginkgo’s lesson is in what it sold.