agent-core

Unified Agent Core — a declarative, provider-agnostic runtime for building LLM agents from YAML manifests. It's the shared core that powers AgentForge and, unmodified, the naturalist assistant in FloraLens — proving one runtime can serve very different products.

Features

• Declarative manifests — define an agent (model, prompt, tools, limits, guardrails, io-schema) in YAML; load + validate fail-fast.
• Pluggable registries — register model providers, tools, prompts, memory backends, and guardrails; extend without editing the core.
• LangGraph ReAct runtime with live token streaming over an SSE-friendly trace bus.
• Model providers — OpenAI, Anthropic, and an offline Echo provider for deterministic tests.
• Built-in tools — web search, HTTP fetch, sandboxed code execution, embedding search; plus MCP auto-binding for external tool servers.
• Multi-agent supervisor — expose sub-agents to a supervisor as tools.
• Guardrails & io-schema enforced at runtime; memory (in-memory / mem0); opt-in durable SQLite checkpointer for multi-turn threads.
• Evaluation harness — dev/held-out suites, programmatic/rubric/LLM-judge scoring, and a regression gate.

Install

pip install "pdz-agent-core[openai]"   # + OpenAI provider; imported as `agent_core`
# extras: [anthropic] [mem0] [mcp] [dev]

Quickstart

from agent_core import build_default_registries, compile_agent, load_manifest_dict

manifest = {
    "id": "echo_agent",
    "model": {"provider": "echo", "name": "test-model"},
    "prompt_ref": "prompts/echo_agent.md",
    "tools": [],
}
agent = compile_agent(load_manifest_dict(manifest), build_default_registries())

async for event in agent.astream("hello"):
    print(event.type, event.detail)

Documentation

Full architecture, API, and the cross-product reuse write-up live in the AgentForge repo under docs/.

License

MIT © pdz1804