Agents
An agent gathers a model, a prompt, and a set of tools under one key — composed with sub-agents, authored in Studio, and versioned in Git. The harness that runs an agent you compose yourself, end to end, is coming soon.
The record
Each agent is one record under a unique key — the handle other surfaces reference it by. Pick the model it runs and the prompt it follows, attach the tools it can call, and drop in a configuration object for anything the runtime needs.
Tools it can call
A tool is a named record with a description and a configuration object, wired to a backend that carries it out — most often a Bosca script you wrote or an MCP server you registered. Register a server by its transport — streamable HTTP, SSE, or stdio — and bind a tool to it.
Author now, Harness coming soon
Agents, prompts, tools, and MCP servers are records you author in Studio and keep in Git, so your AI setup lives beside your code with a full history and a clean way to review a change. Today these parts already power AI across the platform; the harness that runs an agent you compose yourself, end to end, is on the way.
Assemble a model, a prompt, and tools — and sub-agents to delegate to.
Author in Studio and round-trip to Git, with a history you can review.
The harness that runs an agent you compose yourself is coming soon.
Keep exploring