ava-dynamic-workflow

Write an orchestrator script that calls ava.agents.spawn once per sub-task with a self-contained prompt. Each worker writes its result file and terminates silently, while a watcher checkpoint wakes the orchestrator to reduce the results.

Workers write a JSON result file to a shared handoff directory and then terminate themselves. A gather_files watcher polls that directory and sends one message to the orchestrator when the expected files exist, so workers never message directly.

Use it when a task splits into independent sub-tasks, is too large for one agent, or benefits from concurrent execution. Skip it for single-step lookups or strictly sequential tasks, where spawning agents adds overhead without benefit.

Orchestration is a plain Python script instead of a YAML or DAG config, decomposition happens at runtime by the LLM rather than at compile time, and parallelism is true process-level multi-agent execution rather than threads in one process.

Yes. Set REQUIRED_COUNT to K for a K-of-N checkpoint that wakes after any K results land, or use MATCH_GLOB to count files by pattern when result names are unknown. Stragglers keep running and can be terminated later.

Common anti-patterns include workers messaging the orchestrator on completion, arming one checkpoint per worker, polling in a sleep loop, spawning agents for trivial lookups, and over-decomposing into too many workers.