spec-agents-core (agent_fabric)

Domain-agnostic core for spec-driven pipelines and hierarchical agents. Components, pipelines and agents are declared in Markdown with YAML frontmatter (SKILL.md, AGENT.md, fabric.md); the frontmatter is a contract validated by Pydantic and the body is guidance injected into prompts on demand. LLMs plan, delegate, repair and interpret — they never compute: every result comes from deterministic Component.compute() code run by the PipelineExecutor.

Status: pre-release. Until the first PyPI release, install a wheel from the GitHub Releases or use uv sync --all-packages in a checkout; the pip install lines below are the PyPI names.

Names

The distribution is named spec-agents-core because the shorter names are taken on PyPI; the import name and the entry point do not change.

Install

pip install spec-agents-core                  # core: pydantic, numpy, pyyaml, rich (import name: agent_fabric)
pip install "spec-agents-core[tabular]"       # + pandas (dataframe artifact types)
pip install "spec-agents-core[all]"           # + PydanticAI, DSPy, CrewAI, LangChain adapters

Domain packs (statistics, lakehouse, coworker, text-pack) register themselves through the agent_fabric.domains entry point.

Use

agent-fabric catalog --skills path/to/skills     # list components and pipelines
agent-fabric agents path/to/config               # render and validate an agent tree
agent-fabric lint --agents path/to/config --strict
agent-fabric run path/to/config "task" --input name=file.txt

from agent_fabric import build_registry
from agent_fabric.agents import AgentFabric, AgentsConfig

fabric = AgentFabric(build_registry(discover=True), AgentsConfig.load("config"))
report = fabric.run("question", {"data": df}, session_id="s1")

Optional frameworks are imported lazily; models are referenced by LiteLLM alias only. See the repository docs for the architecture, the error model and how to add components, pipelines and agents.