ThinkLess¶

Stop using an LLM for every decision.

ThinkLess is a decision plane for AI agents. The routine judgments an agent makes are answered by rules and small calibrated models in milliseconds; the LLM is kept for the steps that need it and for the decisions the small models are unsure about. Every step is traced with the plane that handled it, its confidence, latency, tokens and cost.

The ThinkLess trace viewer

Where to start¶

New here: Getting started installs ThinkLess and walks from a first decision to a traced cascade.
Evaluating the idea: Benchmarks has the numbers, the method, and what the numbers do not show.
Building an agent: The four planes explains the architecture, and the support agent shows it end to end.
Going to production: Calibration and the production checklist.

In one example¶

from thinkless import Choice, Engine
from thinkless.llm import TransformersLLM
from thinkless.providers import GLiNER, Laya, LLMDecider, Rules

llm = TransformersLLM("Qwen/Qwen3-1.7B")
engine = Engine([Rules(), GLiNER(), Laya(), LLMDecider(llm)], llm=llm, threshold=0.8)

intent = engine.decide(
    "I was charged twice for order #4471",
    Choice("What does the customer want?", options=["refund", "order_status", "other"], name="intent"),
)
print(intent.value, intent.confidence, intent.provider)   # refund 0.99 gliner

The application reads a typed Decision. Which provider answered, how confident it was, and what it cost are in the decision and in the trace.

Design principles¶

Model-neutral. Rules, local encoders, hosted System One models and any LLM are providers behind one interface. The architecture outlives any single model.
Measured, not assumed. Thresholds come from calibration on labeled data; benchmarks ship with their result files and can be rerun with one command.
Safe by default. Uncertain decisions never match is_(). Consequential actions belong behind deterministic checks in code, not behind a model.
Observable. Every decision is a span with its plane, confidence, latency and cost, exportable to OpenTelemetry.
Small core. The base install needs pydantic, httpx, rich and typer. Model libraries are optional extras.