Agently documentation for building AI applications with stable outputs, observable actions, and durable workflows.
Terms are listed alphabetically. Where a term has changed meaning compared with older docs, the entry says so.
The middle layer of Agently’s three-layer action stack: TriggerFlow (top, orchestration) → ActionRuntime (planning + dispatch) → ActionExecutor (atomic backend execution). ActionFlow is the bridge between the runtime and the flow.
ActionRuntime, ActionFlow, and ActionExecutor are now the public plugin types. The older ToolManager plugin type is kept for legacy use only and emits deprecation warnings. See Action Runtime.
Settings on a TriggerFlow execution. With auto_close=True (the default), the execution closes itself after auto_close_timeout seconds of idle. Hidden execution sugar (flow.start() / flow.async_start()) defaults to auto_close_timeout=0.0. flow.start(auto_close=False) is illegal and raises.
The dict returned by execution.close() / execution.async_close(). It captures the final state of the execution. If a compatibility result was written via the deprecated set_result() or .end(), it appears as the "$final_result" key inside the snapshot. See Lifecycle.
In (TypeExpr, "description", True) the third slot is the ensure flag — it marks a leaf as required. Ensure-marked leaves are compiled into ensure_keys (with array wildcards like resources[*].url). YAML / JSON form: $ensure: true.
This is not a default value. The older “default value as third slot” convention is no longer supported, and $default in YAML is gone. See Schema as Prompt.
A single run of a TriggerFlow. Created by flow.create_execution(...). Has lifecycle states open → sealed → closed.
Flow-scoped shared data. Calling get_flow_data(...) / set_flow_data(...) and friends emits a RuntimeWarning because the value is shared across all executions of the flow, which causes problems for concurrency, save/load, and distributed runs. Pass no_warning=True to suppress the warning when the shared scope is intentional. Prefer state for execution-local data.
flow.start() / flow.async_start() create a temporary execution under the hood, run it to close, and return the close snapshot. Convenient for scripts and one-shot runs. Not appropriate for flows that pause for human input, expect external emit()s, or otherwise need an externally-controlled execution handle — use flow.start_execution(...) for those.
The three protocol-level model request plugins: OpenAICompatible, OpenAIResponsesCompatible, and AnthropicCompatible. Most Chat Completions compatible providers configure OpenAICompatible; Responses API-shaped endpoints use OpenAIResponsesCompatible; Claude configures AnthropicCompatible. See Models Overview.
Execution-local storage for live objects — database clients, callbacks, sockets, function pointers, cache handles. Runtime resources are not serializable and do not enter close snapshots or save/load checkpoints; only their resource_keys are recorded. On resume after load(), the caller must re-inject them.
This is a distinct concept from state and flow_data. See State and Resources.
A per-execution stream of items emitted by chunks via data.put_into_stream(...) / data.async_put_into_stream(...). Consumed by execution.get_runtime_stream(...) / execution.get_async_runtime_stream(...). The stream is closed as part of execution.close().
The middle lifecycle state. execution.seal() / execution.async_seal() stops the execution from accepting new external events but lets already-accepted events, internal emit chains, and registered tasks finish. It does not close the runtime stream and does not freeze the close snapshot — that happens on close().
The current name for Agently’s prompt-side structured authoring style: nested dicts of leaves, where each leaf is (TypeExpr, "description", True) and the third slot is ensure. The older “Agently DSL” framing — which tried to be a unified IR for .output(), TriggerFlow contracts, and external schemas — is archived.
Execution-local, serializable, snapshot-safe data. The recommended state surface is data.async_set_state(...) for writes and data.get_state(...) for reads. State is what populates close snapshots and what save() / load() round-trip.
The orchestration layer. Owns branching, concurrency, batches, loops, sub-flows, pause/resume, persistence, and runtime stream. Sits above the Action Runtime and below your application code.
A reserved state key written by the deprecated set_result() and .end() paths. Its presence in a close snapshot indicates a compatibility result was provided. New code should rely on the snapshot itself rather than this key. See TriggerFlow Compatibility.
Deprecated parameter on flow.start(), flow.async_start(), start_execution(), execution.start(), and friends. The value is now ignored with a warning; return shape is controlled by auto_close (and the choice between hidden sugar and explicit execution).