Tool2Agent – a protocol for LLM tool feedback workflows (tool2agent.org)

tool2agent is a lightweight protocol for improving how LLM-driven agents interact with real-world tools by returning rich, structured feedback (errors, suggestions, guardrails) rather than expecting full domain constraints to be encoded in prompts. The idea: many business rules are hidden, dynamic, or too large to fit into an LLM context, so agents should iteratively refine tool calls via trial-and-error guided by machine-readable feedback. That preserves separation of concerns, reduces prompt bloat, and makes agent workflows resilient to changing domain logic. Technically, tool2agent is a set of conventions and developer bindings that make tool-call feedback predictable and programmatically consumable: @tool2agent/ai (AI SDK bindings), @tool2agent/types (TypeScript spec), and @tool2agent/schemas (Zod schema generators that map domain types into tool2agent schemas). Key trade-offs: dynamic, feedback-driven schemas reduce token costs and avoid encoding every tool’s precise schema up-front, but require more tool calls and runtime validation. The protocol also enables reusable middleware and tool-builder utilities to capture common validation patterns as code instead of ad-hoc prompt hacks. It’s an experimental approach, promising for production agent ergonomics and extensibility, and invites builders to explore middleware and tool-building patterns around structured feedback.