Compare · category hygiene
Task models under agents — not the neighbors
Specialist Factory trains small open-weight task models on your proprietary outcomes for one scoped job and puts them under the agent you already run. Not a specialist agent framework, not a RAG platform, not a frontier API — the model under the agent. Shared, BYOC, or confidential GPU.
Vs agent builders
Unlike: LangChain, LlamaIndex, CrewAI, custom harnesses
Sit under those stacks. Your agent calls a task-model endpoint for one labeled job.
Sell orchestration, tools, and agent graphs. They do not replace the frontier call with a measured open-weight model on your outcomes.
Vs fine-tune APIs
Unlike: Together, Fireworks, Predibase, Anyscale FT meters
Recipes + eval vs frontier on your held-out set + trust ladder (Shared → BYOC → Confidential GPU). Product is the task model under the agent.
Sell train/serve infrastructure. Useful meters — different buyer story if you need prove-then-swap and data-path tiers as product.
Vs RAG platforms
Unlike: Vector stores, retrieval-first copilots
Classify or decide on a scoped labeled loop. Keep RAG for open-ended document work; put a task model under the high-volume decision call.
Retrieve context for generation. Wrong tool when you already have ground-truth outcomes and a flaky/expensive frontier classifier.
FAQ
Agent builders sell orchestration. We sit under those stacks: train a task model on your labels and point the existing agent call at it. We are not an alternative agent framework.
Fine-tune APIs sell training/serving meters. We sell measured task models under your agent — recipes, eval vs frontier on your held-out set, and a trust ladder (Shared → BYOC → Confidential GPU).
RAG retrieves context for open-ended answers. Task models classify or decide on one scoped labeled job. Use RAG when you need documents; use a task model when you have outcomes and a decision loop.
See the loop
Definition → how it works → eval methodology → Pilot.