The agent now has a small gearbox.
Strands Decider is a 1.9 billion parameter model that answers typed questions. It can pick one option, answer yes or no, or rate an item against an ordered rubric. The current repository describes a pointer head that scores supplied options in one forward pass instead of generating text token by token.[1]
That shape fits repetitive agent work. A harness can route a ticket, choose a model, check proposed tool arguments, or score an answer without paying for a full language-model turn. The maintainers report a 115 ms median per JevBench question on an RTX 3090 for the v19 architecture. They also warn that the public benchmark has 231 tasks and that small differences between single training runs remain unresolved.[1]
A quick classification can shorten the route. It cannot own the consequence.
Read confidence as a measurement.
The project reports calibrated confidence, but its model card puts a hard boundary around that claim. Calibration uses one temperature per question type, fitted on held-out short classification tasks. The card says to measure on your own traffic before trusting a threshold.[2]
OpenAI's decision-model guide makes the same operational point. The threshold should follow the cost of false positives and false negatives. Lowering the threshold accepts more items and more mistakes. Raising it rejects more items and misses more good ones. The guide recommends tuning on labeled production examples rather than copying a threshold from a demo.[3]
A reported confidence of 0.90 is not a universal 90 percent warranty. It is a score produced by one model, calibration method, input distribution, question, and option set. Change those conditions and the old threshold needs another test.
Keep policy outside the model.
The decision result should enter a policy layer that knows the cost of being wrong. A low-risk routing choice can accept a lower threshold and recover in the next queue. A destructive tool call needs a narrow allowlist, argument checks, a high threshold, and a human or deterministic fallback. A security denial should not disappear because a model sounds certain.
Separate four records. Save the exact model and prompt contract. Save the scores for every option, not only the winner. Save the policy branch that interpreted those scores. Save the action result or escalation. This lets a reviewer tell model behavior from policy behavior.
DECISION FUSE CONTRACT MODEL=StrandsAgents/strands-decider-2B-hobson-v21 QUESTION=Which queue owns this request? OPTIONS=billing,sales,retail THRESHOLD=[TUNED ON LABELED WORKLOAD] BELOW_THRESHOLD=ESCALATE ABOVE_THRESHOLD=ROUTE PROHIBITED=GRANT CREDENTIALS OR DESTRUCTIVE AUTHORITY RECORD=model, revision, input hash, scores, policy branch, outcome
Test the uncertain middle.
The easy examples make a smooth demo. Production lives near the boundary. Build a holdout with ambiguous language, missing context, class imbalance, changed option order, new product names, and costly mistakes. Replay that set after any model, question, rubric, or threshold change.
The repository is unusually frank about current limits. Its model card says changed question wording may leave an answer unchanged, long multi-step documents are weaker, and yes/no or scoring tasks can transfer poorly to new rubrics.[2] Those are test cases, not footnotes.
Installability is real, but the artifact is young. The published strands-decider package was version 0.1.0 when checked. Its CLI exposed ask and serve. Loading a checkpoint still means downloading the base weights and meeting the runtime requirements.[4]
Wire this before the first live call.
DECIDER PRE-FLIGHT 1. Name one bounded decision and all allowed options. 2. Label a workload sample that includes costly errors. 3. Record false-positive and false-negative costs. 4. Tune the threshold on that workload, not a vendor example. 5. Define the below-threshold route and a hard prohibited set. 6. Log every option score, policy branch, and final outcome. 7. Replay the holdout after any contract or model change. 8. Keep irreversible authority outside the score.