bbagents

Python building blocks for agent systems: tools, planning, model selection, memory, human approval, tracing, and evaluation. Works with LangGraph or your own runtime.

Companion code for Agentic AI: Architecture, Theory, and Practice. Buy the book.

Requires Python 3.11+.

pip install -e '.[dev]'
pytest

Complete application example

Agent Stranded is a more complete example: an autonomous survival game using bbagents planning, tools, memory, skills, communication, and tracing, with an optional local LLM. Its code, documentation, and run instructions live in that separate repository.

Features

The previous README's approval/recovery example is preserved as recoverable_release.py.

• Typed tools, dependency-checked plans, structured model decisions and validation.
• Per-agent model selection, routing constraints, cost estimates and caching.
• Capability policies, human intervention, explicit delegation and LangGraph interrupts.
• Bounded async execution, retry policies, execution journals and effect reconciliation.
• Versioned checkpoints, SQLite persistence, task budgets and completion gates.
• Budgeted context, summarization contracts and provider-neutral retrieval (RAG).
• Scoped, revisioned memory and skills; evaluated and approved procedure promotion.
• Roles, messages, delegation and organizational governance.
• Structured traces, isolated evaluation trials and compaction-continuation checks.
• Optional Ollama, LangChain, LangGraph, LangSmith and Langfuse integrations.

No mandatory runtime dependencies. Applications own domain behavior and adapters.

See examples, demos and the guide.

Maintainers: see release instructions for publishing to PyPI.