Agently TriggerFlow

Use TriggerFlow when the application owns visible multi-stage execution semantics. Use ModelRequest/AgentExecution for one bounded request/run, and use agent.create_task(...) when one Agent owns a long task's planning, evidence, verification, and repair without application-authored stage topology.

Read by Need

• Graph construction, state, lifecycle, Stage bridging, and main-repo examples: references/overview.md.
• Pause/resume snapshots, resource restoration, retention, and restart claims: references/recovery.md.
• Optional runtime guidance versus required external wait: references/runtime-intervention.md.
• Stable application stream projection: references/stream-bridge.md.
• DevTools graph/observation: references/devtools-graph.md.
• Full value/signal topology review: ../agently-design/references/execution-topology-validation.md.
• Direct Stage scopes, settlement, bridges, channels, or listeners use agently-stage.

Topology First

Before implementation, map:

• required serial value edges;
• independent branches and bounded joins;
• request-time observation boundaries;
• provisional work that is safe to cancel or discard;
• side-effect ordering, external capacity, waits, repair, and terminal states.

Use batch(...), for_each(...), when(...), and managed emits to make fan-out and joins graph-visible. Represent repetition with an explicit back edge; do not hide lifecycle, retry, or revision loops inside while True chunk handlers.

A ModelRequest is a dispatch-time snapshot. instant output may update UI or start idempotent/cancelable preparation, but a later model request that needs the resulting observation must start after a visible join and validation barrier.

Graph adjacency proves activation, not value transfer. For audits, trace exact values, signals, refs, and consumers through ModelRequest, Action, subflow, TaskWorkspace/RecordStore, wait/resume, repair, and terminal boundaries.

Lifecycle and State

• Prefer async handlers and execution APIs when the caller owns the async boundary. A synchronous provider facade may use Agently-Stage internally; route direct bridge questions to agently-stage.
• Use flow.start(...) / flow.async_start(...) only for finite self-closing runs whose caller needs no execution handle.
• Use flow.create_execution(auto_close=False) for external emit, pause/resume, save/load, intervention, inspection, cancellation, or host-controlled close.
• Start with a positional value and close with await execution.async_close(). Close drains execution-managed nowait work and reports unresolved ownership.
• Put post-resume behavior in a downstream chunk, an explicit resume event, or a data.is_resume branch; a suspended Python frame is not the recovery contract.
• Execution state owns per-run chunk handoff. In async chunks, await async state, emit, and stream methods. Setters replace the complete value; append only for intentional list accumulation.
• flow_data is shared across executions. Save/load serializes and replaces a copy of that shared value; it does not provide isolation, CAS, merge, or concurrency safety.
• TaskWorkspace owns files, RecordStore owns durable records and recovery, and host storage owns business persistence. Keep compact refs/status in execution state and full bodies cold.

Concurrency, Waits, and Streams

• create_execution(concurrency=N) bounds execution-wide handler dispatch; batch(..., concurrency=N) and for_each(..., concurrency=N) bound local fan-out.
• Host admission, provider/model rate limits, and blocking worker pools remain separate pressure owners.
• Use emit_nowait(...) / async_emit_nowait(...) instead of untracked asyncio.create_task(...); execution close settles registered work.
• Use pause_for(..., resume_to=...) for required external input. Use runtime intervention only for optional context at declared boundaries.
• PolicyApproval owns framework policy gates; ExecutionExchange adapts host UI/webhook/queue transport; TriggerFlow owns the interrupt/resume ledger.
• Translate model parser events into stable application events. Do not expose raw parser paths as the frontend protocol, and reconcile provisional items against final validated output.

Recovery and Dynamic-Graph Boundary

Save/load owns TriggerFlow progress and declared recovery metadata, not live clients, callbacks, semaphores, coroutine frames, secrets, or external session state. Restore live ExecutionResources through host/plugin resolvers and verify external refs, versions, leases, and fence tokens before readiness. A local RecordStore proves local restart only; do not claim distributed recovery without a real shared provider and operational evidence.

Define developer-owned stable topology directly in importable TriggerFlow modules with top-level handlers. Explicit submitted or model-generated DAG data is a low-frequency TaskDAG case; read ../agently/references/task-dag.md. Never compile unvalidated runtime plan data into ad hoc TriggerFlow definitions.

API Shape

from agently import Agently, TriggerFlow

flow = TriggerFlow(name="workflow-name")
factory_flow = Agently.create_trigger_flow("factory-workflow")
execution = flow.create_execution(auto_close=False)

when, emit_nowait, and pause_for are flow/runtime methods, not top-level imports. Do not use @flow.when(...) as a decorator or pass a flow name as the first positional TriggerFlow(...) argument.

Anti-Patterns

• A custom event bus, state machine, DAG scheduler, or shadow execution store.
• Sleeps, polling, local completed sets, or untracked tasks in place of signals, joins, and execution-managed work.
• Closure-captured live business resources instead of explicit runtime resources and resolvers.
• flow_data as ordinary per-execution state.
• DevTools diagrams as topology source of truth instead of the definition and runtime metadata.