Agentic coding harnesses, catalogued
Your team has been asked to adopt agentic engineering. Coding agents such as Claude Code, Cursor, and Codex can write code, but they bring no process of their own: whether they plan first, write tests, get reviewed, or record what they learned depends on the person prompting them. A harness is a layer you install into the agent that supplies that process, as skills, commands, subagents, and hooks. There are many of them, and they make different trade-offs. This site describes how each one works, lists its features in the same format, and gives you a process for choosing.
Start here
- How harnesses work: the parts every harness is built from, and the terms this site uses.
- How to choose a harness: the questions to answer before you pick one.
- Browse harnesses by category, or compare their features side by side.
Categories
- Workflow methodologies (3) - An opinionated development process, from idea to merged code, packaged as skills, commands, and subagents.
- Spec-driven toolkits (0) - Turn requirements into a written spec and task list before any code is written.
- Multi-agent orchestrators (0) - Run and coordinate several agents at once, usually in separate sessions or worktrees.
- Autonomous loops (0) - Keep an agent working toward a goal over many iterations with little supervision.
- Memory and context (0) - Give agents persistent memory or curated context across sessions.
- Skill collections (0) - Libraries of independent skills, agents, or commands with no single prescribed process.
Recently researched
- Compound Engineering (2026-09-28) - draft
- pstack (2026-09-28) - draft
- Superpowers (2026-09-28) - draft