Agents#
The framework provides three core concrete agent entry points. All three use
the same public delegate contract: compile(prompt, ...) for workflow
construction and run(prompt, ...) for execution. Choose by execution
pattern first, then by control requirements.
Multi-agent orchestration patterns (PlanExecutePattern,
ProposeCriticPattern, TreeSearchPattern, RalphLoopPattern, RouterDelegatePattern) live in the patterns module and are
implemented with the same public workflow step primitives available to users.
SeededRandomBaselineAgent is a lightweight benchmarking/control-condition
participant for packaged-problem studies. It follows the same workflow-backed
runtime shape as the other public agents; supply packaged-problem objects
through the run-time dependencies mapping.
PromptWorkflowAgent wraps a prompt-mode Workflow for packaged-problem
studies. Use it when the workflow is the real participant implementation but
the experiment loop should own problem resolution, run ids, and condition
selection.
Overview#
DirectLLMCallMultiStepAgent(mode="direct" | "json" | "code")SeededRandomBaselineAgentPromptWorkflowAgent
Decision table#
Use case |
Recommended pattern |
Why |
|---|---|---|
Plain text generation without tools |
|
Lowest orchestration overhead |
Iterative direct (no external tools) |
|
Internal CONTINUE/STOP controller steps |
Iterative structured tool loops |
|
ReAct-style multi-step loop with JSON |
Iterative code-action loops |
|
ReAct-style loop with code actions |
Packaged-problem control condition |
|
Thin seeded baseline over public problem contracts |
Packaged-problem prompt-mode workflow participant |
|
Reuses a public |
Background references#
These references are for conceptual grounding; behavior is defined by the contracts and implementation in this repository.
Examples#
examples/agents/README.md