WorkflowSkill: A Standardized Workflow Language Specification for AI Agents

WorkflowSkill is an open-source workflow language specification designed to enable AI agents to delegate predictable tasks to deterministic runtimes for execution, achieving durable execution, automatic retries, and cross-platform portability, thus providing engineering reliability guarantees for AI application development. Its core philosophy is "Agents improvise, workflows execute reliably", representing a crucial step towards engineering and industrialization in AI application development.

Current Large Language Model (LLM) agents excel at reasoning and creative tasks, but face fundamental challenges when executing structured, multi-step, repetitive tasks: having the agent "improvise" to complete the same task flow each time not only consumes expensive tokens and causes reasoning delays, but also carries risks of inconsistent results or even hallucinations. The WorkflowSkill project was born to address this pain point.

The project's slogan "Agents improvise. Workflows deliver." accurately summarizes its design philosophy: agents are good at reasoning, decision-making, and creative work, while workflows excel at deterministic, repeatable, and observable execution. Combining the two enables: reduced reasoning costs (only intelligent steps call LLM), improved reliability (automatic retries, timeout control, breakpoint resumption, etc.), ensured consistency (same input produces same output), and enhanced observability (step input/output records facilitate debugging and auditing).

WorkflowSkill adopts a modular dual-component architecture: Toolkit is responsible for routing workflow actions to actual APIs/SDKs, with the core method execute(action, args, idempotencyKey) -> unknown; the built-in Weldable simulation tool includes 11 common integrations (e.g., Anthropic, Discord, GitHub, etc.). Runtime is responsible for orchestration and execution, providing two implementations: in-memory (for local development and testing) and the planned dbos (for persistence and crash recovery). The separated design ensures that the same workflow YAML can run cross-platform.

Workflows are defined using YAML, including metadata (version, name, inputs/outputs, etc.) and steps. There are various step types: action (call external integration), transform (JSONata data conversion), if/switch (conditional branching), foreach (iteration), while (loop), parallel (parallel execution), try/catch/finally (error handling), wait/wait_for_signal (waiting). The expression language uses JSONata, and context variables include step outputs, input parameters, workflow metadata, etc.

Reliability features for production environments: automatic checkpoints (created after step execution, supporting crash recovery), idempotency guarantees (automatically generated keys to avoid side effects of repeated execution), retry mechanisms (configurable max_attempts, backoff strategies, error code whitelists, etc.).

Suitable scenarios: scheduled automation (data synchronization, report generation), event-driven processing (responding to webhooks/system events), multi-system integration (coordinating SaaS services), long-running processes (approval, supply chain coordination). Comparison with existing solutions: declarative YAML definition, AI-native, specification-driven cross-platform, optional persistence, open-source—differentiated from Zapier (visual locked platform), Temporal (code-based), Airflow (Python code), etc.

Currently in the early development stage, it has implemented a YAML specification interpreter, simulation toolkit, in-memory runtime, and Claude Code integration; planned features include DBOS runtime, more integrations, a visual editor, and a workflow marketplace. Conclusion: WorkflowSkill promotes the transformation of AI applications from "prompt engineering" to "software engineering", providing reliable infrastructure for production-level AI applications, and is worth developers' attention and participation.