Skip to content
Tembo Agent Studio Tembo Agent Studio

Core concepts

A few ideas explain how everything in TAS fits together.

Agents are code

An agent is a file in your repo under agents/, not a record in a vendor console. Its model, instructions, connections, and tools are all declared in that file. Because it’s a file, it has a history, a diff, an owner, and a review trail — the same as any other code.

Every change is a pull request

You never edit a live agent in place. Changes — new agents, edits, and “improve the agent” feedback — are turned into a pull request against your repo. The authoring work (chat → diff) is performed by the Tembo Coding Agent Platform using the API key you set in workspace settings. You review and merge the PR like any other; the merged file is what runs.

Workspaces

A workspace maps to one GitHub repository and owns its members, connections, agents, automations, and runs. Switching workspaces switches all of that. Access is governed by roles.

Two agent frameworks

TAS runs two agent file formats, each as a passthrough into the upstream tool so you get its full power:

  • Pydantic AgentSpec — the canonical authoring format. A YAML (or JSON) file with model, instructions, optional connections:, and TAS extensions like labels: and tools_module:. Runs through a bundled pydantic-ai wrapper.
  • Cargo AI — a single-file JSON agent format, importable and run through the bundled cargo-ai CLI.

The Agents list and run pipeline detect the framework from the file’s shape.

Draft → stable versioning

An agent has a draft (the live file on your default branch) and, once you promote it, a stable snapshot frozen in the database. Runs default to the stable snapshot for predictability; the chat/iterate surface always runs the draft. Promotion records who promoted what and when, and you can compare versions — so “what changed and when did behavior shift?” is always answerable.

Connections and tools

Agents reach the outside world through Connections (authorized per user) and the tools those connections expose. TAS has two connection substrates — Composio and Native MCP — plus the option of sidecar Python tools for deterministic work. See Connections and Tools & Tool uses.

Acting user

Every run executes as a user, and that identity decides which connection credentials it uses. A manual run acts as you; a scheduled run acts as the automation’s owner; an event-triggered run acts as the trigger’s owner. This is why connections are authorized per person.