da-factory

Dynamic Agent Factory.

Declarative agent manifests -> runnable Strands agents.

Hand the SDK a manifest (prompt, tools, guardrails, context policy) and get back a correctly-wired, runnable agent -- without hand-assembling hooks/interventions boilerplate per agent, per project.

No dependency on strands.multiagent.* (Graph/Workflow): orchestration is concurrent (fan-out + optional aggregator) or sequential (chain), both built on nothing but strands.Agent.invoke_async + asyncio.

Status

Early alpha. API surface is still settling.

Install

pip install da-factory

Quick example

from da_factory.catalog.discovery import discover_agents
from da_factory.catalog.build import build_agent
from da_factory.orchestration import run_concurrent

manifests = discover_agents("path/to/definitions")
agents = {m.id: build_agent(m) for m in manifests}
result = await run_concurrent(agents, "some input")

CLI

New project: create a definitions/ directory, then scaffold agents into it with new-agent (no separate "init project" command -- the directory itself is the project).

# LLM-backed agent (prompt + tools, no logic.py)
da-factory new-agent my-agent --role "role description" --llm

# deterministic/code-backed agent (logic.py, no prompt)
da-factory new-agent my-agent --role "role description"

# target a non-default definitions directory
da-factory new-agent my-agent --role "role description" --llm --definitions-dir path/to/definitions

Validate all manifests under a definitions directory:

da-factory validate --definitions-dir path/to/definitions

Guardrails

Built in, declared per-manifest via guardrails: [injection, pii, secrets, toxicity]. Each follows the same two-layer shape: a deterministic pre-filter, plus (where it adds value) a semantic LLM judge invoked only when the pre-filter finds nothing.

Development

uv sync --all-extras --dev
uv run pytest
uv run ruff check . && uv run ruff format --check .
uv run pyright

License

See LICENSE -- not yet finalized (placeholder, no license granted).