What is agent-ftw?

agent-ftw is a TypeScript framework for building AI agents that work in structured, sequential phases rather than as a single open-ended loop. The core idea is that you define an agent as a series of phases, each with its own prompt, a JSON schema describing exactly what output it must produce, and an optional turn budget (a cap on how many back-and-forth exchanges the model gets before the phase must either succeed or fail). When a phase finishes, only its structured output carries forward to the next phase, the raw conversation history is discarded, so the context doesn't balloon over a long pipeline. This schema-enforced handoff between phases is the main differentiator the README emphasizes. Because each phase is required to emit output that matches a defined structure before it can end, the next phase always receives clean, typed data rather than hoping the model remembered to include a field. You can also route different phases to different model backends: a paid frontier model for the phases where quality matters, and a free local model (via Ollama, LM Studio, or vLLM) for cheaper verification steps. An optional checklist feature runs a second model as a quality judge on a phase's output before accepting it. Sub-agents and scoped detours ("side quests") are also supported, each still ending with a typed deliverable. It works with any OpenAI-compatible endpoint and with the Claude SDK, so you can mix providers across phases. The runtime has minimal dependencies: it requires Node.js 22 or newer and a TypeScript toolchain, and the Anthropic SDK adapters are optional and installed separately only if you need them.