Agently Dynamic Task

Use this skill when the problem is a dynamic task graph: a model or application submits DAG data that must be planned, validated, pruned, resolved to handlers, and executed.

TaskDAG is the Agently DAG foundation capability. Dynamic Task is the current compatibility and convenience facade over TaskDAG planning and execution, not a TriggerFlow sub-API and not the strategic Agent task lifecycle. TriggerFlow is the execution substrate under TaskDAGExecutor, while Agently.create_dynamic_task(...) and agent.create_dynamic_task(...) remain compact app-facing facade entrypoints.

Native-First Rules

• use Agently.create_dynamic_task(...) or agent.create_dynamic_task(...) when ordinary app code wants the compact facade
• when a DAG is only one bounded phase inside a broader business task, keep agent.create_task(...) or agent.create_task_loop(...) as the outer AgentExecution owner and expose the DAG as an explicitly enabled Agent candidate; do not turn the Dynamic Task facade into the long-running AgentTask lifecycle owner
• split into AgentlyTaskDAGPlanner, TaskDAGValidator, DynamicTaskResolver, and TaskDAGExecutor only when staged control is required
• use plan=<TaskDAG> when the caller already owns the DAG and should skip model planning
• use TaskDAG.from_yaml(...) or TaskDAG.from_json(...) when the submitted DAG is a reviewed config artifact
• use submitted DAG inputs placeholders for small runtime wiring: ${INIT} / ${INIT.foo} for initial graph input, ${DEPS.task_id.path} for dependency results, ${STATE...} for execution-local state, and ${TRIGGER...} for the raw TriggerFlow trigger payload. Slot names are case-insensitive, but examples should use uppercase. Whole-string placeholders preserve the original value type; embedded placeholders stringify into the surrounding text. Missing runtime paths fail closed during execution
• when a submitted DAG runs through agent.create_execution(), let ${INIT...} read use_dynamic_task(graph_input=...) when explicit, otherwise the frozen execution prompt input slot, then {"target": task_target}; do not add a second inputs.input or task-local prompt mapping surface
• use planner=<agent-or-provider-settings> when the model must generate the DAG
• use model=<agent-or-provider-settings> for model task execution resources
• use handlers={"risk_check_handler": handler} for local/custom tasks; handler names should be explicit and usually end in _handler
• pass actions=... or skills=... only when the planner should be allowed to use those task kinds; do not expose them by default
• use Agently output schemas for model-owned semantic outputs instead of parsing model text with example-local JSON/code-block parsers
• keep backend business-system calls mockable in examples, but keep model-owned planning, classification, drafting, evaluation, revising, and final response generation model-driven
• require each model task that matters to have a clear output schema, either through facade-level output_schema or node-level inputs.output_schema
• set model-task inputs.output_format by result type when the plan is generated or submitted: omit it to let the request read prompt.default_output_format (global default json); use json for compact machine-control data, action arguments, routing flags, standalone booleans/numbers, strict extraction, model judges, and dense all-typed arrays/objects; xml_field for flat string long text/code/HTML/SVG/Markdown/SQL/templates; explicit hybrid for long prose/code fields mixed with typed list/object/boolean/number fields when retry latency is acceptable; explicit flat_markdown only for legacy section-header compatibility; explicit yaml_literal only when a YAML target document is intentionally desired; auto only when structural schema-driven selection and retry latency are acceptable after the target model has passed representative stability checks. qwen2.5:7b checks found hybrid can omit section headers or echo old scaffold comments into text fields
• let TaskDAGValidator reject duplicate ids, missing dependencies, cycles, unsupported required task kinds, schema-version mismatches, unsafe side-effect policy, and unknown required handlers before execution
• allow unknown optional handlers to be pruned only when they do not affect required semantic outputs, required downstream nodes, approvals, or side-effect policy
• keep generated plan data in execution input or execution state; do not use shared TriggerFlow flow data for per-run DAG state
• use ${INIT...} / ${DEPS...} placeholders for small declarative input wiring; for larger transformations, keep the logic in a handler or model task
• do not teach planners to use internal resolver entries such as validate or emit as ordinary task kinds
• do not mark ordinary read-only model tasks as network side effects just because they call an LLM provider
• do not add approval gates to analysis, synthesis, drafting, or final response model tasks unless the user explicitly asks for approval

API Shape

Submitted DAG:

from agently import Agently

async def risk_check_handler(context):
    return {
        "task_id": context.task.id,
        "deps": dict(context.dependency_results),
    }

task = Agently.create_dynamic_task(
    target="review policy",
    plan={
        "graph_id": "review",
        "task_schema_version": "task_dag/v1",
        "tasks": [
            {"id": "extract", "kind": "local", "binding": "risk_check_handler"},
            {
                "id": "final",
                "kind": "local",
                "binding": "risk_check_handler",
                "depends_on": ["extract"],
                "inputs": {
                    "source": "${INIT}",
                    "extract_result": "${DEPS.extract}",
                },
            },
        ],
        "semantic_outputs": {"final": "final"},
    },
    handlers={"risk_check_handler": risk_check_handler},
)

snapshot = await task.async_start(timeout=10)
result = snapshot["semantic_outputs"]["final"]

Config-backed submitted DAG:

from agently import Agently
from agently.core import TaskDAG

graph = TaskDAG.from_yaml("plans/review.yaml")
task = Agently.create_dynamic_task(
    target="review policy",
    plan=graph,
    handlers={"risk_check_handler": risk_check_handler},
)
snapshot = await task.async_run(graph_input={"doc": "policy"}, timeout=10)

Use TaskDAG.from_json(...) for JSON/JSON5 files or raw content. Both loaders support task_dag_key_path="plans.review" for selecting one DAG from a larger config. Use graph.get_yaml(path) or graph.get_json(path) to export a normalized graph.

Auto-planned DAG:

from agently import Agently

task = Agently.create_dynamic_task(
    target=user_goal,
    planner=planner_agent,
    model=worker_agent,
    output_schema={
        "answer": (str, "final user-facing answer", True),
    },
)

graph = await task.async_plan(max_retries=3)
task.validate(graph, strict_schema_version=True)
snapshot = await task.async_run(graph_input={"goal": user_goal}, timeout=30)

Lower-level integration:

from agently.builtins.plugins import AgentlyTaskDAGPlanner
from agently.core import DynamicTaskResolver, TaskDAGExecutor, TaskDAGValidator

resolver = DynamicTaskResolver({"risk_check_handler": risk_check_handler})
validator = TaskDAGValidator(resolver)
planner = AgentlyTaskDAGPlanner(validator=validator)

graph = await planner.async_plan(planner_agent, {"target": "review policy"})
validation = validator.validate(graph, strict_schema_version=True)
snapshot = await TaskDAGExecutor(resolver, validator=validator).async_run(graph)

Planner Constraints

Planner plugins and planner prompts must declare:

• task schema version, normally task_dag/v1
• allowed task kinds and available binding names
• required semantic output leaves
• maximum task count when the scenario has a practical limit
• validation checks that become retryable model-output failures
• forbidden behavior: unstable task ids, dependency results embedded in static task inputs, implicit side effects, hidden action/skill use, and internal resolver task kinds

The facade planner should wire .output(planner.output_schema(), format="json"), required planner.ensure_keys(), and .validate(planner.validate_output) so structural problems become model-output validation failures before execution. Compile/run still revalidates the graph.

Layer ownership:

• TaskDAG / TaskDAGNode data contracts and YAML/JSON graph config belong to agently.types.data
• TaskDAGExecutor, TaskDAGValidator, and DynamicTaskResolver belong to core
• AgentlyTaskDAGPlanner owns planner output schema, ensure keys, and planner instructions as a plugin concern
• Agently.create_dynamic_task(...) and agent.create_dynamic_task(...) are the app-facing facade entrypoints

Anti-Patterns

• do not put Dynamic Task guidance under TriggerFlow as if it were TriggerFlow syntax
• do not hand-roll a scheduler for submitted DAG data
• do not parse model text with json.loads(...), regex marker splits, or code-block extraction when Agently output schemas can own the shape
• do not expose actions or skills to the planner by default
• do not hide model-owned classification, drafting, quality checks, or final response generation behind deterministic local substitutes in recommended examples
• do not call low-level TaskDAGExecutor from ordinary app code when Agently.create_dynamic_task(...) is enough
• do not use vague handler names such as risk_check; use risk_check_handler or another name that describes the callable role
• do not duplicate TriggerFlow docs here; explain only that Dynamic Task compiles to TriggerFlow internally

Read Next

• references/overview.md
• examples/minimal.py