langchain-loadout 0.2.2

Per-turn skill selection for deepagents. Load the instructions each turn needs, instead of the full catalog.

Quick start · Results · Documentation · Contributing

Why Loadout

Large skill catalogs take up context on every model call. Loadout ranks and verifies skills for each user turn, then loads the relevant instructions or suggests candidates. The agent can also search the catalog with find_skill.

Loadout extends deepagents through LangChain middleware. Its routing core can also be used independently of either framework.

Quick start

1. Install. Requires Python 3.11+. The jev extra includes the judge adapter and its SDK.

pip install "langchain-loadout[jev]"

2. Connect your skills. Place them in ./skills/<name>/SKILL.md with name and description in YAML front matter. Set TYPESAFE_API_KEY and your model provider's credentials (ANTHROPIC_API_KEY for this example).

from deepagents import create_deep_agent
from deepagents.backends import FilesystemBackend

from langchain_loadout.langchain import LoadoutSkillsMiddleware
from langchain_loadout.providers.jev import JevJudge

backend = FilesystemBackend(root_dir=".", virtual_mode=True)
loadout = LoadoutSkillsMiddleware(backend=backend, sources=["/skills/"], judge=JevJudge())

agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-5",
    backend=backend,
    skills=["/skills/"],
    middleware=[loadout],
)
await agent.ainvoke({"messages": [{"role": "user", "content": "I need a statement for the embassy"}]})

Run the example in an async context. Selection runs on ainvoke and astream; synchronous calls use the ordinary skills middleware. Tune thresholds and questions through Settings.

The public API is evolving: minor releases may introduce breaking changes while the version is 0.x.

Results

Latest benchmark on a bank-statement assistant using deepagents and langchain-loadout 0.1.0 from PyPI: 236 skills, 50 conversations × 5 turns per variant (250 turns each).

The skills section was 17.2× smaller, with 3.3× fewer input tokens, but cost per turn was 44% higher. Smaller prompts do not necessarily mean lower cost when cache reuse changes.

These results use generated data, one judge and one agent model; the two-point accuracy difference alone does not establish an accuracy improvement. Fit thresholds to your own data. Earlier measurements and design trade-offs →

Documentation

License

MIT © 2026 Ivan Deyna