Tip: Choose your execution strategy before wiring agents together

Pick the right ExecutionStrategy subclass up front — swapping strategies later requires rethinking data flow between every agent in the pipeline.

Why: Each strategy defines not just execution order but also how results pass between stages, so the wrong choice compounds across every step.

Use get_strategy(strategy_name) to retrieve a strategy by name rather than instantiating subclasses directly. If the name isn't registered, you get an explicit ValueError listing available options — far cheaper than a silent misconfiguration discovered mid-pipeline.

from attune.orchestration import get_strategy

strategy = get_strategy("sequential")  # raises ValueError with available names if unknown

For pipelines that reuse the same context across many agents, prefer PromptCachedSequentialStrategy over a plain sequential approach — it shares a cached context with a configurable TTL (cache_ttl, default 3600 s), which avoids redundant computation on each agent call.

For hierarchical delegation, DelegationChainStrategy enforces a max_depth (default 3) that prevents runaway recursive delegation. If you find yourself raising that limit, treat it as a signal that the pipeline needs restructuring rather than a deeper chain.

Tradeoff: get_strategy() returns a pre-configured instance. If you need non-default parameters — such as a custom max_depth for DelegationChainStrategy or a specific cached_context string — instantiate the class directly instead of going through the registry.

Source files

Tags: orchestration, teams