Pattern: Universal Tool Wrappers With AGENT.md Discovery Give Agents Deterministic, Zero-Token Access to Any Interface

GitHub·medium signal

OpenCLI's AGENT.md pattern standardizes how AI agents discover and invoke tools — each wrapped interface publishes a machine-readable manifest describing available commands, input schemas, and output formats. This is converging with Claude Code's CLAUDE.md and similar conventions: agents need a deterministic, documented surface to operate reliably. The zero-LLM-token-at-runtime claim is the key insight — agent tool invocation should be structured, not prompt-interpreted.