Orchestration

Orchestration is the system that selects an execution strategy, assembles the right agents, and routes data between them so that complex multi-step tasks run as a single coordinated pipeline.

Mental model

Every orchestrated task moves through three layers:

Strategy selection — MetaOrchestrator examines the task's complexity, domain, and requirements, then picks a composition pattern (Sequential, Parallel, Debate, Teaching, Refinement, or Adaptive). You can also call get_strategy(name) directly to bypass automatic selection.
Agent assembly — The registry supplies AgentTemplate instances matched by capability or tier. Dynamic teams can be built at runtime via DynamicTeamBuilder and stored in TeamStore.
Execution — The chosen ExecutionStrategy subclass calls execute(agents, context), enforces any quality gates or depth limits, and returns a StrategyResult that downstream stages or callers consume.

This separation means you can swap strategies — for example, replacing SequentialStrategy with ParallelStrategy — without changing the agents or the code that reads the result.

Execution strategies

Each strategy implements ExecutionStrategy.execute(agents, context) -> StrategyResult. The strategies available out of the box are:

Strategy	What it does
`SequentialStrategy`	Runs agents one after another, passing output forward
`ParallelStrategy`	Runs agents concurrently and merges results
`DebateStrategy`	Agents argue positions; a judge resolves conflicts
`TeachingStrategy`	A lead agent instructs subordinate agents
`RefinementStrategy`	Each agent iteratively improves the previous output
`AdaptiveStrategy`	Switches sub-strategy at runtime based on intermediate results
`ToolEnhancedStrategy`	Single agent with access to the full tool set
`PromptCachedSequentialStrategy`	Sequential execution with a shared cached context (default TTL: 3600 s)
`DelegationChainStrategy`	Hierarchical delegation enforcing a configurable `max_depth` (default: 3)
`ConditionalStrategy`	Branches execution: if condition then A else B
`MultiConditionalStrategy`	Switch/case branching across multiple conditions with an optional default
`NestedStrategy`	Embeds a registered `WorkflowDefinition` as a single step
`NestedSequentialStrategy`	Sequential steps where any step can be an agent or a nested workflow

Use register_strategy(name, strategy_class) to add your own implementation to the registry.

Workflow composition

Workflows are named, reusable pipelines registered with register_workflow(workflow) and retrieved by ID with get_workflow(workflow_id). NestedStrategy and NestedSequentialStrategy reference them by WorkflowReference, which lets one workflow embed another up to a configurable nesting depth.

A concrete example from the meta-orchestration patterns: a Secure Release pipeline uses SequentialStrategy to chain security_audit → dependency_check → release_prep, where each stage is itself a registered workflow.

Agent registry

The registry decouples strategy logic from agent definitions:

get_template(template_id) — retrieve one template by ID
get_templates_by_capability(capability) — find agents that expose a specific capability
get_templates_by_tier(tier) — find agents preferring a resource tier
register_custom_template(template) — add a user-defined agent at runtime
unregister_template(template_id) — remove a template; returns False if the ID is not found

Key interfaces

Other parts of the codebase interact with orchestration through these entry points:

Interface	Purpose	Source
`ExecutionStrategy`	Base class every strategy extends	`_strategies/base.py`
`ToolEnhancedStrategy`	Single-agent execution with full tool access	`_strategies/advanced_strategies.py`
`PromptCachedSequentialStrategy`	Sequential execution with shared prompt cache	`_strategies/advanced_strategies.py`
`DelegationChainStrategy`	Hierarchical delegation with depth enforcement	`_strategies/advanced_strategies.py`
`ConditionalStrategy`	If/else branching between workflow branches	`_strategies/conditional_strategies.py`
`MultiConditionalStrategy`	Switch/case branching with optional default	`_strategies/conditional_strategies.py`
`NestedStrategy`	Embed a registered workflow as a single step	`_strategies/conditional_strategies.py`