deerwork-ai/deer-workflow

What it solves

Deer Workflow provides a way to build AI agent graphs that are observable and reusable. It solves the problem of "opaque" agent conversations by moving the control flow, failure handling, and execution paths into reviewable TypeScript code, ensuring that the plan is explicit rather than hidden within a model's internal logic.

How it works

The system uses a "code-first" approach to graph engineering. Developers define the execution paths in TypeScript, while coding agents (such as Codex, Claude Code, or Pi) perform the actual semantic work within each node of the graph. It includes a workflow-creator skill that allows users to describe a desired orchestration in plain English, which the system then converts into a runnable TypeScript module.

Who it’s for

It is designed for developers who want to build complex agentic workflows with precise control over the execution flow and the ability to swap out different AI agent runtimes without changing the core logic.

Highlights

• Code-as-Plan: Control flow and inputs are defined in TypeScript for better reviewability.
• Vendor-Neutral Agents: Supports multiple agent runtimes including Codex (default), Claude Code, and Pi.
• Observable Execution: Features a phase-aware TUI for interactive runs and a JSONL event stream for automation and CI/CD pipelines.
• Automated Generation: Can generate complete workflow modules from natural language descriptions.