Agent
Reason, explore, inspect, code, delegate, and decide while the path is uncertain.
Flows
Turn repeatable agent work into durable visual flows.
Let agents explore and design. Move dependable execution into an explicit engine with queues, retries, checkpoints, backpressure, and recovery.
The division of labor
Reason with agents. Run with flows.
Flow Editor
Turn the discovered procedure into tasks, connections, queues, conditions, schedules, and subflows.
Runtime Viewer
Observe execution, queue pressure, task state, checkpoints, and recovery as the Flow runs.
Deterministic core
Make reliability visible.
Author in the right viewUse technical processors or declarative steps while one canonical Flow and its named layouts remain the source of truth.
Review before executionA planner can propose an exact draft revision; acceptance and approval create one immutable version and durable one-shot run.
Pause for peopleTyped questions, confirmations, timers, waits, and notifications park durably and resume from the same interaction in Web, PawCode, or VS Code.
Persist and recoverQueues, checkpoints, bounded loops, retries, and backpressure keep progress explicit across interruptions.
AI inside a Flow
Use models as tasks, not as the scheduler.
A Flow can call an agent or LLM exactly where judgment is useful while deterministic tasks own collection, transformation, persistence, retries, and delivery.
Workflow AgentsOpt in to bind an agent to an exact Flow version, frozen resources, finite limits, durable checkpoints, and inspectable recovery.
Bounded deliberationCompose named agents in a reviewed group workflow without merging their tool permissions or giving the group ambient authority.
Human-visible executionStart and stop tasks, inspect queues and run events, and understand the current state without reading an opaque transcript.
Apache NiFi meets Claude Code
Build with reasoning. Operate with structure.
Start with a planning request, review the proposed Flow, then run the approved version with durable state.