STEP is an agent-readable protocol for durable, resumable work across long-running agents, humans, and tools. It preserves intent, execution state, and evidence so work can survive context loss, support handoff, and continue across execution sessions.
A STEP WorkNode is executable intent with persistent state and evidence:
PLAN.md— intended workSTATE.md— current execution statelogs/— execution evidence
This repository publishes a software idea, not a software package.
applications/ contains non-normative mappings for long-running agents, deep research, and human-agent handoff.
PROTOCOL.md defines STEP and is authoritative. examples/hello-worknode/ is a minimal, non-normative example. provenance/worktree-bootstrap/ preserves the historical WorkNode that directed the early WorkTree implementation.
There is no canonical implementation. Implementations are contextual, replaceable expressions of the protocol. They may use any language, agent, toolchain, or isolation strategy that preserves STEP's semantics.
The first STEP protocol interpreter was generated using STEP. Originally exposed as worktree, it deterministically selects actionable work and manages WorkNode state. It remains the kernel beneath the author's current step application. The implementation is evidence of the protocol in use, not part of the protocol itself.
Read PROTOCOL.md. From it, a human or agent should be able to understand the execution model, determine what work is actionable, resume or hand off work, validate outcomes from recorded evidence, and implement an appropriate runner.
The protocol is the durable software. Implementations can be generated again.