spec-agents

spec-agents (the agent-fabric toolkit) lets you build pipelines and agent teams from Markdown specs, with one rule above all: LLMs plan, delegate, repair and interpret; code computes. A model never produces a result by itself. Every number, table and file comes from a deterministic component that a spec declares and the toolkit validates before it runs.

What you get

• Specs in Markdown. A component, pipeline or agent is a SKILL.md / AGENT.md file: the YAML frontmatter is the contract (ports, params, budgets), the body is guidance for the model.
• Verified before it runs. Plans are validated against the registry and report all problems at once as typed, located errors.
• Bounded agents. Depth, delegations and LLM calls are budgeted per run, so a team cannot loop forever.
• Offline-first. Everything in this guide runs without a model; a scripted backend plays the LLM.
• Domain packs. The core knows no domain; packs add components and pipelines for statistics, lakehouse engineering, code review and text.

Where to start

I want to…	Go to
install it and run something in five minutes	Installation then Quickstart
use a domain pack on my own data	Domain packs
understand a failure	Handle errors
point it at a real model	Use a real model
look up an API, or contribute	API reference and contributor guides

The code on these pages is not pasted: each block is a script from examples/docs/ that the test suite runs offline, so what you see works.