AgentML – HTML for Agents (Deterministic, MIT-Licensed) (www.agentml.dev)

AgentML is a new MIT‑licensed, XML‑style language and runtime for defining deterministic agent behavior — pitched as “HTML for agents.” Instead of relying on brittle natural‑language prompts, developers write structured AML (AgentML) documents that describe states, transitions and typed data models. The project includes a CLI and AMLX runtime that interprets those documents, plus a quick install option, with the goal of turning prompt engineering into system design: author once, validate with W3C‑compliant schemas, and run anywhere with full traceability. Technically, AgentML claims 100% conformance with W3C SCXML (all 193 tests), full XSD 1.0 support with automatic import/include resolution and schema caching, and built‑in identity constraints and type checking. It logs every step and tool call for replay/debugging, offers validators/guards to enforce safety and policy before execution, and supports thread‑safe concurrent validation/execution. The runtime integrates with Model Context Protocol (MCP) tool servers and typed LLM calls, letting teams combine deterministic state machines with modern model backends. For AI/ML teams this promises reproducible, auditable, and safer agent behavior—reducing nondeterminism from prompts while preserving LLM integration and production throughput.