Skip to main content

​
Agent

A named model with a system prompt, a model role and a list of tools it may call, defined in an agent file. Agents run in graph ask, graph chat and the workbench, and plans call them as agent__<name> tools. See Agents.

​
Agent network

Any composition of agents, plans and tools that call one another. See Core concepts.

​
Agent step

The agent control step, which hands one plan step to a model that calls tools until it can return a result. An agent step is defined inline in the plan. See agent.

​
Ask step

The ask control step, which asks the person running the plan a question and uses the answer. See ask.

​
Body

The steps a route branch, map or reduce runs: a single call or an inline list of steps. See Conditions and bodies.

​
Built-in

An agent, plan or tool that ships with graph. See Customize a built-in agent.

​
Catalog

Every tool available to an agent or plan, each named with a prefix that says where it comes from, such as user__ or plan__. See Core concepts.

​
Control step

A step that steers how a plan runs instead of calling a tool. It branches (route), narrows a list (filter), repeats over a list (map, reduce) or stops the plan (exit), or hands the step to a model (agent) or a person (ask). See Control steps.

​
Decision model

A model that answers a fixed question with a likelihood or a choice among options, instead of writing text. Gates using decide and decision tools use one through the decider model role. See Models and providers.

​
File version

The schema version a config, plan, tool or agent file carries in its version key, separate from graph’s release version. Each graph release reads a range of file versions and upgrades older files as they load. See File versions.

​
Finish

How a plan ends and what it returns: silent, an output map built from step results, or a solver that writes an answer. See finish.

​
Gate

The condition on an exit, route or filter step: a comparison of values, an infer yes-or-no verdict from a model, or a decide call to a decision model. See Conditions and bodies.

​
Handoff

One agent passing the conversation to another through a transfer_to_<name> tool. The other agent answers from then on. See Build a multi-agent network.

​
Layer

One of the three places graph loads agents, plans, tools and config from. Built-ins load first, then your global files in ~/.config/graph/, then project files in ./.graph/. A file in a later layer replaces one with the same name in an earlier layer. See Core concepts.

​
Model role

A name such as chat, judge or a role you define, which agents, steps and tools use instead of naming a model. [models] in config.toml maps each role to a model. See models.

​
Pack

A set of tools, plans and agents installed from a git repository with graph pack add, such as github or slack. Each pack is a [packs.<name>] entry in config, and its tools are called as <name>__tool__<tool>. See Packs.

​
Plan

A file of steps wired together by their results. Each step’s input draws on earlier results through templates, and control steps branch, repeat and stop the flow. See Plans.

​
Shape

The observed output schema of a tool, with an example, recorded each time the tool runs. See Improve what graph knows about a tool.

​
Step

One entry in a plan: a tool call, a control step, or a call to an agent or another plan. See steps.

​
Subagent

An agent another agent calls as a tool. The subagent runs, returns a result, and the caller keeps the conversation. See Build a multi-agent network.

​
Template

A {{...}} reference in a step, tool or prompt that splices in an input or an earlier result. See Template language.

​
Thread

The saved history of a conversation with an agent in graph ask or graph chat. Continue one with --thread. See Manage threads.

​
Tool

Anything an agent or step can call: MCP server tools, your tool files, built-in and pack tools, and plans and agents exposed as tools. See Tools.

​
Tool kind

What a tool file runs: exec runs a command, prompt calls a model, reshape rebuilds JSON without a model, and decision asks a decision model. See Tool file.

​
Workbench

The full-screen terminal app opened by graph workbench, where you draft, run and debug plans, agents and tools. See Workbench.