Concepts overview
The mental model behind kdeps: how a workflow is shaped, how data moves between steps, and which pieces apply to each mode. If you have not run kdeps yet, read What is kdeps? first.
How a workflow is shaped
A kdeps project is a folder. workflow.yaml is the manifest - name, version, and targetActionId (the resource that produces the final result). Every other .yaml file under resources/ is a single resource: one LLM call, one SQL query, one shell command, one HTTP request. A resource lists what it requires:, and kdeps runs the resources in that dependency order, passing each one's output forward. That ordered graph is the workflow.
The tables below group every concept by what it does. Start with the two modes, then reach for the rest as you need them.
The two modes
| Concept | What it is | Mode |
|---|---|---|
| Workflow mode | A DAG of resources that runs in dependency order and returns a response | Workflow |
| Agent mode | An LLM that calls workflows, components, and built-in tools to finish a task | Agent |
| Agencies | Several agents bundled under one agency.yaml, calling each other with agent: | Both |
| Components | Reusable resource bundles you install or build, invoked with component: | Both |
Wiring data between steps
| Concept | What it is | Mode |
|---|---|---|
| Expressions | expr-lang snippets in before:/after:, and check: | Both |
| Expression helpers | Json(), Safe(), Debug(), default() and other utilities | Both |
| Data access | get()/set() plus the input and request shorthands | Both (input/request: workflow) |
| Tools (function calling) | Let a chat: resource call other resources mid-response | Workflow |
| Inline resources | chat/sql/python/... actions nested in a resource's before:/after: | Both |
Control flow and input
| Concept | What it is | Mode |
|---|---|---|
| Validation and control | The validations: block - skip, preflight check, route/method limits, input schema | Workflow |
| Error handling (onError) | Retry, fallback value, or fail on a resource error | Workflow |
| Items iteration | Run a resource once per list entry (for-each) | Both |
| While-loop | Repeat a resource while a condition holds; every: for scheduled tasks | Both |
| Input sources | Feed a workflow from an HTTP API, a chat bot, or a file | Workflow |
| Jinja2 templates | YAML preprocessing and project scaffolding | Both |
State
| Concept | What it is | Mode |
|---|---|---|
| Session storage | Values that persist across requests from the same caller | Workflow |
| Persistent memory | Project-scoped facts the agent recalls across sessions | Agent (tools also work in workflow) |
| Memory internals | Auto-extraction, the memory graph, and prompt injection | Agent |