Sequential Executor

Why it matters

A fixed pipeline has no model to repair a failed step, so how the step fails is decided in configuration. This example makes that decision explicit.

What to practise

Install the example and run it with nat run: a sequential_executor workflow whose tool_list passes text through a processor, an analyzer and a report generator, each output becoming the next input, with no agent or LLM involved. Then try the three ways a step can end the chain. raise_type_incompatibility decides whether a mismatch between adjacent type annotations, checked before the chain runs, raises or only warns. return_error_on_exception either re-raises a tool’s exception or ends the chain early and returns an error message as the output. A tool can raise SequentialExecutorExit to stop the chain on purpose with a message of its own. tool_execution_config sets use_streaming per tool.

How it fits

These settings are 2.4.1’s routing choice in miniature: a caller exception, an error returned as output, or a deliberate early outcome. When a step has side effects, ask 2.4.4’s question of what an early exit leaves half done.