How bingreeky/opd-evolver works
OPD-Evolver is a slow–fast agent-evolution framework. Its runtime workflow runs benchmark tasks through an agent, retrieves and selects hierarchical experience for execution, then reflects on results to update memory and collect training data. A separate slow loop prepares OPD data and trains policy capabilities. The graph retains the documented benchmark and memory-management branches; several runner and training internals are unsampled, so their connections are shown only where the README or source index supports them.
Main entry points: task_runner.py Task runners · memory_pipeline.py Memory pipeline · react_agent.py ReAct agent · memory_manager.py Hierarchy manager
Components 19
Agent Runtime
- Task runners
opd_evolver/runners/task_runner.py - ReAct agent
opd_evolver/subagents/react_agent.py - Step memory
opd_evolver/base/agent/memory.py - Main agent
opd_evolver/main_agent.py - Benchmark evaluation
scripts/eval/bench_simple_awm.py
Experience Memory
- Memory pipeline
opd_evolver/pipelines/memory_pipeline.py - Selection and reflection
opd_evolver/pipelines/memory_prompts.py - Hierarchy manager
opd_evolver/memory/memory_manager.py - Experience tiers
opd_evolver/memory - Embedding search and storage
opd_evolver/memory/base_store.py - Embedding providers
opd_evolver/memory/embeddings.py - Selector and writer data
opd_evolver/memory/selector_dataset.py
Policy Learning
- Executor OPD dataset
scripts/dataset/build_awm_executor_opd_dataset.py - OPD trainer
opd_evolver/trainer/opsd_trainer.py - Deployable policy
Model Services
- LLM client
opd_evolver/base/engine/async_llm.py
Other components
- Task initiator
- Benchmark tasks
- Embedding service
Connections 22
- Task initiator → Task runners: starts task
README.md - Benchmark tasks → Task runners: supplies tasks
README.md - Task runners → ReAct agent: runs agent
opd_evolver/runners/task_runner.py - ReAct agent → Benchmark tasks: acts in environment
opd_evolver/subagents/react_agent.py - ReAct agent → Step memory: records steps
opd_evolver/base/agent/memory.py - Task runners → Memory pipeline: invokes memory flow
opd_evolver/runners/task_runner.py - Memory pipeline → Hierarchy manager: retrieves and updates
opd_evolver/pipelines/memory_pipeline.py - Memory pipeline → Selection and reflection: uses prompts
opd_evolver/pipelines/memory_pipeline.py - Memory pipeline → Selector and writer data: logs examples
opd_evolver/pipelines/memory_pipeline.py - Hierarchy manager → Experience tiers: manages tiers
opd_evolver/memory/memory_manager.py - Hierarchy manager → Embedding search and storage: delegates search
opd_evolver/memory/memory_manager.py - Embedding search and storage → Embedding providers: uses embeddings
opd_evolver/memory/memory_manager.py - Embedding providers → Embedding service: requests vectors
- Memory pipeline → LLM client: generates reflections
opd_evolver/pipelines/memory_pipeline.py - ReAct agent → LLM client: requests actions
- Task runners → Executor OPD dataset: rollout data used to build
README.md - Selector and writer data → OPD trainer: supports capability training
README.md - Executor OPD dataset → OPD trainer: trains on dataset
README.md - OPD trainer → Deployable policy: updates policy
README.md - Deployable policy → Benchmark evaluation: evaluated by
README.md - Task runners → Benchmark evaluation: benchmark rollout
README.md - Main agent → LLM client: uses model client
opd_evolver/main_agent.py