playbooks

Playbooks is a full semantic programming system for AI agents: a structured Markdown surface language for specifying agent behavior, a semantic intermediate representation called PBAsm, a compiler from Markdown to PBAsm, and a durable runtime that executes PBAsm with the determinism of traditional software and the flexibility of an LLM. Agents are modeled as classes with public playbooks (methods), triggers (event decorators), instance state, and a real call stack. It shipped on PyPI as playbooks through 16 releases ending at v0.7.4, spawned an org-scale ecosystem of sibling projects (VSCode debugger, fine-tuned LM, durable runtime extracted as OSS, docs site, enterprise edition), and is the result of three years of sustained solo work on a single thesis: natural language as a programming language.

It is a complete system that took the idea seriously - PBAsm as an IR, class-based agent modeling with triggers, a VSCode debugger for natural-language programs in May 2025 well ahead of the industry, and an explicit forward-compatibility principle. Active development is paused rather than finished. A published retrospective works through what the simpler generalist-agent-plus-skills approach got right, and is equally clear about which capabilities here - a real call stack for natural-language programs, step debugging, a compiler that turns prose into a verifiable instruction set - still have no equivalent anywhere else. The code, the releases, and the documentation all remain available, and parts of it are candidates for revival.