Execution modes
Sequential pipelines and orchestrator coordination in Floprints. Two execution paradigms from one WorkflowPlan schema.
Two modes, one schema
Every WorkflowPlan declares an execution strategy as a discriminated union:
Sequential
A fixed linear pipeline: step_1 → step_2 → … → step_N.
- Each step maps inputs from
$.userInput.<field>or a prior step’s output ($.steps.step_K.output[.field]) - Deterministic and auditable — the same inputs follow the same path
- Best for regulated, repeatable processes
Orchestrator
A coordinator agent decides at runtime:
- Which agents to call and in what order
- When to run independent agents in parallel
- When to pause and ask the user a question
- When the workflow is complete or should fail
Best for open-ended tasks where the path cannot be fully specified upfront.
Run engine behavior
The execute service drives both modes:
- Initialize run — persist inputs, create step records
- Invoke harnesses — call Bedrock AgentCore with resolved tools
- Record messages — every turn with token usage and latency
- Handle human pauses — set phase to
waiting_human, surface questions - Resume — replay human answers into the same harness session
Continuing orchestrator runs
Finished orchestrator runs are continuable — follow-up messages append to the persistent session so multi-turn workflows stay coherent.
Execution overrides
Run requests can pass overrides (bounded by deployment caps) to adjust iteration limits or timeouts for a single invocation without changing the deployed plan.