Course outline for Dynamic and Agentic Workflows

Module 1: Treat generated plans as proposals

• Deterministic vs stochastic stages: Separate stable transformations from variable planning proposals so generated decisions cannot bypass fixed execution and acceptance controls.
• Plan generation: Express report plans as typed proposals with declared stage inputs, dependencies, and actions rather than immediately executable instructions.
• Plan validation: Check dependencies, input compatibility, and action policy independently of JSON validity; reject cycles and unauthorized writes.

Practical outcome: Define typed plan constraints and reject schema-valid proposals with invalid dependencies or unauthorized writes.

Module 2: Compile accepted dependencies

• Dynamic DAG construction: Build an acyclic execution graph only from accepted stages, preserving typed artifact dependencies and validated action boundaries.
• Agent stages: Bound variable planning work with explicit input and output contracts while keeping stable operations in the local runner.
• Execution plan compilation: Translate validated proposals into executable stages and bind recorded approval to the plan hash and input version.

Practical outcome: Compile validated stages into a local runner and record approval against the plan hash and input version.

Module 3: Adapt without bypassing acceptance

• Workflow adaptation: Use failure feedback to propose limited plan revisions, comparing the added complexity with a simpler fixed workflow.
• Runtime decision-making: Check whether changed inputs or plans invalidate accepted execution and stop before stale approval authorizes changed actions.
• Human intervention: Request renewed approval for revised plans and inputs, preserving versioned acceptance records before execution can resume.

Practical outcome: Use failure feedback to propose a bounded revision, stop stale execution, and obtain renewed acceptance before changed actions.

Practical exercise

• Produce typed plan fixtures, a validator, compilation diagnostics, and versioned execution records for a synthetic report task.
• Accept a valid plan; reject cycles, incompatible stage inputs, and an unauthorized action even when JSON is valid.
• Execute stable operations in the local runner and verify dependency order and the mock output artifact.
• Change an input or plan after approval; verify that execution stops until renewed approval, then compare the design with a fixed workflow.