Core Concepts
The mental model behind taskflow's declarative DAG runtime.
This chapter explains the stable execution model beneath every taskflow. It does not enumerate every field — use the Syntax Reference for that — and it does not cover the 0.2 compiler surfaces — use New in 0.2.
A taskflow rests on six ideas:
- work is a graph, not an implied sequence;
- twelve phase types build the graph;
- data moves through explicit interpolation;
- the graph is verified before token spend;
- intermediate output and filesystem work can be isolated;
- persisted state lets a run resume across sessions.
Suggested reading order
The DAG Model — dependsOn creates edges; array order does not.
Phase Types — the twelve building blocks and when to use them.
Interpolation — how values move between phases without an extra model call.
Verification — cycles, dead ends, references, and budget checks before execution.
Context Isolation and Workspace Isolation — keep transcripts and file changes inside controlled boundaries.
Shared Context Tree — coordinate isolated agents without dumping every transcript into the host.
Resume — continue incomplete work and reuse persisted phase results.
Replay and incremental recompute build on this model, but are documented as 0.2 runtime capabilities in New in 0.2.
The concepts
The DAG Model
How task nodes and explicit dependencies form the graph.
Phase Types
The twelve execution and control building blocks.
Interpolation
Deterministic data flow between phases.
Verification
Catch structural errors before tokens are spent.
Context Isolation
Why only the selected final result reaches the host conversation.
Workspace Isolation
Run file-changing work in inherited, temp, dedicated, or worktree scopes.
Shared Context Tree
Coordinate agents through explicit shared context.
Resume
Continue persisted runs across sessions.
Last updated on
Was this helpful?
Help us improve the docs or ask a question in the community.