Agents

An agent in LabFrame is not a stateless chatbot. It is an AI researcher with its own dedicated computer. Each agent has:

• A dedicated cloud VM — an isolated Linux machine with its own filesystem, shell, and installed packages
• Persistent storage — files survive across sessions; nothing resets when the conversation ends
• Autonomous execution — the agent can search the web, write code, run commands, and produce deliverables on its own machine
• Configurable behavior — system prompts and model selection shape how the agent works

​

Creating an agent

Click + New Agent in the sidebar. You will see a modal with two fields: After you click Create, LabFrame provisions a dedicated cloud VM. The sidebar shows status updates while the machine starts. When the green dot appears, the agent’s computer is ready and the agent can start working.

​

Machine types

Each agent’s computer has a specific hardware profile. Click the CPU icon in the header to open the machine type selector. Options range from lightweight VMs (2 vCPUs, 4 GiB) to GPU machines (A100s) for heavy workloads. Changing the machine type restarts the agent’s VM. Files on the persistent workspace disk are preserved.

​

Projects

Each agent can have multiple projects — separate conversations, each with its own chat history and optional workspace scope. To create a new project, click the + icon next to the agent name in the sidebar. You can give the project a name, optionally upload starter files, and invite collaborators. Projects keep different tasks organized. A single “Research Assistant” agent might have one project for a literature review and another for data analysis, each with their own conversation thread — but sharing the same underlying computer and workspace.

​

System prompts

The system prompt tells the agent how to behave. You can set it when creating the agent or edit it later. For example:

You are a research assistant for a materials science lab. When asked to review a paper, produce a structured summary with sections for methodology, key findings, limitations, and relevance to our lithium-ion battery degradation project.

Agents with clear, specific system prompts produce better results. Think of the system prompt as the agent’s job description.

​

Models

Use the model picker at the bottom of the chat input to switch models:

• Opus 4.6 — most capable, best for complex multi-step research tasks
• Sonnet 4.6 — fast and capable, good for most tasks
• GPT 5.4 — OpenAI model, strong at coding and structured output

You can switch models between messages. The agent keeps the same computer, workspace, and conversation regardless of which model powers it.

​

Agent states

If your agent is asleep, it wakes automatically when you send a message. You can also put a running agent to sleep using the Moon icon in the header to save credits when you’re not using it.