These examples are the current recommended examples for AgentExecution, Dynamic Task DAG, ActionRuntime, and direct AgentExecution Skill binding.

For changes from released 4.1.4.7, see the bilingual 4.1.4.8 example change guide.

Older Skills auto-orchestration examples from before the 4.1.3.8 Blocks lifecycle refactor were moved to:

examples/archived/pre-4.1.3.8-skills-orchestration/agent_auto_orchestration/

Those archived files are reference material only. They should not be treated as runnable examples or recommended usage for 4.1.3.8 and later, and should not be forced onto the new Blocks lifecycle.

Run from the repository root. Earlier model examples need DEEPSEEK_API_KEY in the environment or .env; set DYNAMIC_TASK_MODEL_PROVIDER=ollama for local Ollama where supported. Examples 25-28 use local Ollama/Qwen directly and default to qwen; override it with AGENT_EXECUTION_OLLAMA_MODEL or OLLAMA_DEFAULT_MODEL.

python examples/agent_auto_orchestration/02_actions_dag_streaming.py
python examples/agent_auto_orchestration/05_model_field_delta_streaming.py
python examples/agent_auto_orchestration/06_parallel_dag_field_streaming.py
python examples/agent_auto_orchestration/20_agent_execution_lineage_workspace_loop.py
python examples/agent_auto_orchestration/21_agent_execution_github_issue_intake.py
python examples/agent_auto_orchestration/22_unified_agent_execution_result.py
python examples/agent_auto_orchestration/23_agent_execution_auto_dispatch.py
python examples/agent_auto_orchestration/24_independent_dynamic_task_dag.py
python examples/agent_auto_orchestration/25_agent_execution_delivery_review_ollama.py
python examples/agent_auto_orchestration/26_plan_execution_interaction_ollama.py
python examples/agent_auto_orchestration/27_long_content_execution_artifact_ollama.py
python examples/agent_auto_orchestration/28_missing_goal_preparation_ollama.py
python examples/agent_auto_orchestration/29_execution_controls_ollama.py
python examples/agent_auto_orchestration/29_field_long_content_ollama.py

_TEMPLATE_standard_skill_orchestration.py shows the released run_skills_task(...) convenience adapter. New code should prefer agent.use_skills(...).input(...) and consume the ordinary AgentExecution.

• 02 - Customer Support Triage. Independent Dynamic Task DAG with local handlers, dependency edges, and real model calls over mocked CRM data.

• 05 - Operator-visible Field Delta Streaming. Independent Dynamic Task DAG with kind="model" nodes and field-level runtime streaming.

• 06 - Parallel DAG Field Delta Streaming. Independent multi-branch Dynamic Task DAG with concurrent workstreams and a fan-in executive brief.

• 20 - AgentExecution Lineage Context Loop. Two-step AgentExecution lineage, RecordStore persistence, and TaskContext disclosure example.

• 21 - GitHub Issue Intake. AgentExecution plus restricted shell Action for real GitHub CLI issue intake.

• 22 - Unified AgentExecution Result. Minimal quick prompt plus task-loop strategy consumed through the same result/stream/meta facade.

• 23 - AgentExecution Auto Dispatch. Route-selection example proving default model_request and task-strategy agent_task dispatch.

• 24 - Independent Dynamic Task DAG. Infrastructure smoke for direct Agently.create_dynamic_task(...) submitted-DAG execution.

• 25 - AgentExecution Delivery And Review. Local Qwen business result, verified TaskWorkspace artifact, model-backed advisory review, and a host-owned blocking review handler.

• 26 - Plan Execution With Interaction. Local Qwen readiness analysis, request-local connected clarification, host-validated structured plan, and verified artifact delivery.

• 27 - Long-Content Execution Delivery. Local Qwen section planning and writing, host-ordered Markdown assembly, verified artifact delivery, and model-backed advisory review.

• 28 - Missing Goal Preparation. An explicitly selected long-task producer asks the model to derive missing goal/criteria from the original request. No Actions are authorized. The early run timed out after preparation; the later release-candidate run completed preparation and accepted production.

• 29 - Execution Controls. Safe outer pause, snapshot/rebind/resume, same-object revision rework, retained readers and explicit cleanup. Request rework preserves original information and instructions alongside the latest feedback.

• 29 - Field Long Content. Explicit LongContent fields use the chapter producer and fill their final strings back into a structured result; conditional continuation remains an independent option.

Early 26/27 runs exposed semantic-quality problems; they are historical observations, not the latest acceptance status. In the final candidate check, 26 preserved its original prompt: local output replaced the explicit metric, while three unchanged-prompt DeepSeek runs preserved the metric and passed the 120-minute plan and artifact checks. Example 27 delivered the artifact and truthfully returned a failed advisory review under the default warn policy; this is not a strict quality guarantee. Model and input quality remain relevant.

Model calls are real. Business data is mocked unless the example explicitly states that it uses a real external system such as MCP or GitHub CLI.