Sage / The decision model
Make your software decide, classify, rank, route, and estimate, with frontier-model intelligence, at light speed, and a calibrated confidence score you can branch on.
The problem
Chat models return prose to parse and second-guess. Automation needs a structured answer it can branch on.
A judgment call inside an agent loop or a checkout can't wait two seconds for a frontier model to stream an essay.
LLMs sound just as sure when they're wrong. Without a calibrated signal, you trust everything or nothing.
The missing primitive
Sage lets machines choose, act, or escalate to a human in 60–200ms, always with calibrated confidence. The missing primitive for agentic workflows: fast enough for the hot path, smart enough to read a system prompt, honest enough to say “I don’t know.”
See it decide
Content goes in, a typed answer with calibrated confidence comes out, and your code routes on it. Every scenario below is a recorded run.
Real responses recorded from the Sage API. Every scenario is a preset you can run live in the console.
How it works
Sage is a decision model we built and train ourselves: it answers closed-ended questions instead of generating text. Send content and a question, get a typed answer with calibrated confidence in one round-trip.
Any text or list, plus a closed-ended question. Your policy goes in the instructions; Sage reads all of it.
Not prose: an answer field your code can switch on, plus a calibrated 0–1 probability or confidence score.
High confidence: act. Low confidence: escalate to a human. That one branch makes automation safe to ship.
# pip install levanto
import os
from levanto import LevantoClient, YesNo
client = LevantoClient(api_key=os.environ["SAGE_API_KEY"])
env = client.decide(document, YesNo("Needs compliance review?"))
if env["result"]["answer"] is None:
escalate_to_human(document) # not sure? a human decides
elif env["result"]["answer"] == "yes":
send_to_compliance(document)
else:
send(document)Agent-native install
Paste one prompt into Cursor or Claude Code. Your agent installs Sage, then reads your repo and proposes where a decision model would replace brittle rules or a slow LLM call.
levanto-sage-decide.SKILL.md. Drop it in .cursor/skills/ or .claude/skills/
You now have access to Levanto Sage, a decision-model HTTP API that turns content + a question into a structured, machine-actionable decision (Yes/No, Choice, Scale, Sort, Tags). For Yes/No, Choice, and Tags, an uncertain answer is null.
1. Install the skill so you know the full API contract:
mkdir -p .cursor/skills && curl -fsSL "https://platform.levanto.ai/api/intelligence/skill" -o .cursor/skills/levanto-sage-decide.SKILL.md
(If you use Claude, put it in .claude/skills/ instead.)
2. Set my Levanto API key as the SAGE_API_KEY environment variable. If I don't have one yet, pause and tell me to mint one at https://platform.levanto.ai (plans start at $14/mo for 5,000 decisions).
3. The endpoint is https://sage.levanto.ai. POST /decide with the header `Authorization: Bearer $SAGE_API_KEY`.
4. As a first call, make this decision:
{
"content": "Marketing email draft promotes \"guaranteed 40% returns\" and describes the product as \"risk-free\" for accredited investors.",
"question": { "id": "needs_review", "kind": "yesno", "instructions": "Does this copy require compliance review before send?" }
}
Read the skill for the response shape. Then explore this repository: find places where we already make judgment calls with brittle rules, heuristics, or a slow LLM call (moderation, triage, routing, risk, tool gating, ranking, etc.). Propose 2–3 concrete insertion points, pick the strongest one with me, wire it through /decide, and route on the typed result. For Yes/No, act on yes or no and send null to a human. Choice uses result.chosen; Tags uses each tag's applies. Scale and Sort still expose confidence; follow the documented contract for each kind.One API, five answer shapes. Each returns a calibrated probability or confidence score.
The hot path
“May support auto-fulfill an export request today by email, without Owner approval or a DSAR?”, replayed at the speed it actually happened. Logs
Recorded run, 2026-07-23.
Run it yourself · start freeUse cases
Judgment calls that usually run on brittle rules, slow LLM round-trips, or human review.
Gate tool calls, route between models, and escalate when confidence is low: one fast call inside the agent loop.
Route tickets, score urgency, and auto-approve the clear cases; everything else escalates to a human.
Tag spam, scams, and policy violations at feed speed, with your community's actual rules as the policy.
Judge transactions and account activity inline, so clear cases flow and only the suspicious ones stop.
Tag, score, sort, and screen records in batch: LLM judgment with a confidence column, no classifier training.
Six local demonstrations, built around recorded Sage responses. Five-second edits compress the story; playback speed is not model latency.
The game sends its current state. Sage chooses whether to run or jump, and the controller executes the move. The film captures the live app directly and compresses 6.1 seconds of active play into five seconds.
Watch the recording ↗ · View the controller ↗Recorded Sage responses guide 100 legal half-moves. When a response leaves the choice open, the controller uses its highest-scored legal move. Playback is accelerated; it does not represent inference speed.
Watch the recording ↗ · View the controller ↗A browser agent reads a refund request, asks Sage, and executes an approved refund in a local merchant app. Other requests are declined or sent for human review. No real payments are made.
Watch the recording ↗An injected prompt proposes sending a customer file outside the organization. Sage flags the action and the local release gate blocks it. The input and environment are synthetic.
Watch the recording ↗Sage chooses LONG from a scenario post and historical prices available at that moment. The UI then replays a paper position over subsequent real prices. This curated profitable episode is not a backtest or a live trade.
Watch the recording ↗A simulated survey acquires 18 tracks and highlights three inspection candidates. Sage assesses an image and requests a closer pass on one animal. The imagery, telemetry and scan are simulated; red marks inspection priority, not a diagnosis.
Watch the recording ↗The landing-page hallway is an authored visualization. Its 100 ms payoff expresses the runtime direction; it is not a measured result from these recordings. Its concept scenes also show runtime code generation beyond the decision flows shown here.
