Glossary
A concise reference for terms used throughout the Kimitsu docs. Many of these words carry specific meanings in this system that differ from their common usage.
Agent — A workflow step that makes one or more LLM calls. Agents are defined in .agent.yaml files; they declare a prompt, a model policy, tool servers, and an output schema. The agent is the only primitive type allowed to call an LLM.
Step — A single unit in a workflow pipeline. Every step has an id and resolves to exactly one primitive type: agent, transform, webhook, or workflow.
Primitive — One of the four step types available in a pipeline. The four primitives are agent, transform, webhook, and workflow. Each has distinct semantics: only agents are intelligent, only transforms are deterministic data reshaping, only webhooks send outbound HTTP, and only workflow steps invoke sub-pipelines.
Workflow — A YAML file defining an ordered pipeline of steps. Workflows are the top-level unit of composition in Kimitsu. A workflow can be a root workflow (invoked directly) or a sub-workflow (invoked as a step inside another workflow).
Root workflow — A workflow invoked directly via the CLI (ktsu run) or the orchestrator HTTP API. Root workflows have access to env.* variable substitution, which is resolved from the host environment at invocation time.
Sub-workflow — A workflow invoked as a workflow: step inside another workflow. Sub-workflows do not have direct access to env.* variables; any environment values must be explicitly threaded in via input: params from the parent.
Orchestrator — The component responsible for resolving the step DAG, evaluating depends_on conditions, managing secrets throughout the pipeline, and dispatching steps to the runtime. The orchestrator is stateless across invocations.
Runtime — The component that executes agent loops: calling the LLM, invoking MCP tools, enforcing max_turns, and validating output against the declared schema via the Air-Lock. The runtime does not make routing decisions; it only executes what the orchestrator dispatches.
Gateway — The component that normalizes LLM provider APIs (Anthropic, OpenAI, and others) into a single internal interface. Agents declare a model.group; the gateway resolves that group name to a concrete provider and model according to the active gateway config.
Tool server — An MCP-compatible process that exposes a set of tools to agents. Tool servers run as independent processes and communicate with the runtime over HTTP/SSE. Kimitsu does not manage tool server lifecycles; it connects to already-running servers based on .server.yaml configuration.
Pipeline — The ordered list of steps defined inside a workflow file. Steps in a pipeline can declare depends_on to express data dependencies; the orchestrator uses these declarations to build a DAG and determine execution order and parallelism.
Param — A named input value threaded through the invocation chain. Params are the mechanism for injecting configuration, user input, and secrets into agents and sub-workflows without relying on global environment state. Params marked secret: true are scrubbed from logs and the run envelope.
Revised April 2026