Problem: AI agents run multi-step workflows — code review, data pipelines, incident response — but there is no standard way to define what they should do or record what they actually did. You get a chat log at best. Good luck auditing that.
OSOP is two YAML formats:
.osop— .osop — workflow definition. Nodes (16 types: agent, tool, CLI, API, human gate, etc.) + edges (13 modes: sequential, parallel, conditional, fallback, spawn, etc.). Think of it as a DAG for agent behavior..osoplog— .osoplog — execution record. What actually happened: timestamps, durations, token counts, tool calls, inputs/outputs, error states. One log per run.
Together they form a closed loop: define, execute, record, optimize.
Quick example (.osop)
Demo: osop-editor.vercel.app — osop-editor.vercel.app — drag and drop any .osop file to visualize it. You get a node graph, risk analysis, step-by-step execution replay (if you load the .osoplog too), and sub-agent tracking for multi-agent workflows.
Why not X?
- BPMN — BPMN — designed for enterprise business processes in 2004. XML-based, massive spec, no concept of LLM calls or token budgets.
- OpenTelemetry — OpenTelemetry — infrastructure-level observability (spans, traces). Does not define what a workflow should do, only instruments what happened at the system level.
- Oracle Agent Specification — Oracle Agent Specification — YAML-based agent definition, similar space. But no execution records (.osoplog), no built-in risk analysis, no human-in-the-loop gates, no self-optimization loop.
- GitHub Actions / Temporal / Prefect — GitHub Actions / Temporal / Prefect — great workflow engines, but proprietary formats coupled to their runtimes. OSOP is runtime-agnostic.
EU AI Act angle
Article 19 requires "automatically generated logs" for high-risk AI systems. .osoplog is structured, machine-readable, and captures exactly what regulators need — model used, inputs, outputs, human oversight points. Compliance as a byproduct.
What exists today
- Spec v1.0 (JSON Schema validated)
- Visual editor (React 19, live at osop-editor.vercel.app)
- Python CLI (pip install osop) — validate, render, run
- JS/TS SDK and Python SDK
- GitHub Action for CI validation
- 39+ example workflows
- 18 platform integrations (Cursor, Codex, Windsurf, Claude Code, Copilot, and more)
All open source, Apache 2.0.
GitHub: github.com/Archie0125
Ask HN: What workflow formats do you use for AI agents? We have been looking at the landscape and most teams seem to either use proprietary formats tied to their orchestration framework, or just wing it with unstructured logs. Curious what others have tried.