Orchestrate complex
workflows with DAGs
Switch any workflow to DAG execution mode. Steps run concurrently based on dependencies, convergence loops refine output iteratively, and SubWorkflowStep lets you compose workflows within workflows.
DAG Execution
Build workflows as directed acyclic graphs
Switch any workflow to DAG execution mode. Steps declare dependencies and the engine executes them in concurrent waves, automatically resolving the optimal order. Use SubWorkflowStep to invoke other workflows as steps, giving you composable multi-stage orchestration from a single canvas.
- Visual DAG canvas with topological layout and Bezier edge rendering
- Concurrent wave execution via dependency graph traversal
- SubWorkflowStep to compose workflows within workflows
- Fan-out (one-to-many) and fan-in (many-to-one) patterns
Convergence Loops
Iterative refinement with quality gates
Link an Eval step to an upstream target. When quality scores fall below your threshold, the executor automatically re-runs the subgraph with eval feedback injected into LLM prompts. Iterations continue until quality is met or the max iteration count is reached.
- EvalStep quality gates with configurable scoring criteria
- Automatic re-execution of target-to-eval subgraph
- Feedback injection into LLM and recipe prompts on each iteration
- Configurable max iterations (1-10) with back-edge visualization
Pattern Templates
Start with proven orchestration patterns
Choose from four pre-built multi-agent orchestration patterns. Each template generates workflow steps, input schemas, dependencies, and convergence loops from your customization parameters.
- Critic/Refiner — LLM + Eval with convergence loop
- Specialist Router — Router step with specialist branch routes
- Debate/Consensus — Parallel LLMs + Aggregator for consensus
- Plan/Execute/Verify — 3-step chain with convergence loop
Crash Recovery
Resilient execution with checkpointing
Enhanced checkpointing tracks per-step completion, output snapshots, and convergence state. If a long-running workflow fails mid-execution, it resumes from the last successful step — in both sequential and DAG modes.
- Per-step completion tracking with output snapshots
- Resume from any point after crash or timeout
- Convergence state preserved across restarts
- Backward-compatible with v1 checkpoint format
Step Types
Five new step types for multi-agent workflows
In addition to recipe, condition, loop, parallel, and approval steps, workflows now support five AI-native step types.
LLM Step
Issue a direct LLM prompt without a recipe wrapper. Ideal for lightweight transformations, classifications, and quick decisions within a workflow.
Eval Step
Score outputs against quality criteria with LLM-as-judge. Trigger convergence loops for iterative refinement until thresholds are met.
Router Step
Classify inputs using an LLM and branch to specialist routes. Each route executes a different downstream path based on the classification.
Aggregator Step
Combine outputs from parallel branches via merge (union), best-of (select highest quality), or synthesize (LLM-generated summary) strategies.
Coding Agent Step
Execute a sandboxed coding agent that writes, tests, and iterates on code within Docker containers. Ideal for data transformations, report generation, and programmatic tasks within workflows.
Plans
Orchestration for every team
DAG execution and all ten step types are available on Pro. Advanced orchestration features are available on Enterprise.
Pro
- Unlimited workflows with all 10 step types
- DAG execution mode with visual canvas
- Workflow version control
Enterprise
- Everything in Pro
- SubWorkflowStep for cross-workflow orchestration
- Canary rollouts for workflow versions
- Convergence loops with crash-recovery checkpointing
Ready for advanced workflow orchestration?
DAG execution and all step types are available on Pro. Talk to our team about Enterprise features.