Your LLM gave your agent language.
Cortex gives it a forecaster.

Cortex is the predictive layer you embed inside your AI agent. It turns your client's first-party data into a twin your agent can query over a stateless API, then runs one loop: predict what's about to happen, simulate the change before it ships, and monitor the outcome to prove the action worked. Every prediction comes back with a confidence interval; every outcome feeds the next model. This page walks the whole thing, end to end.


Predict, simulate, then monitor, one loop your agent runs.

Cortex turns your client's data into a twin your agent can query over the API: simulate a what-if (blue), act on the forecast (jade), monitor outcomes and drift (amber), then learn and predict again. Telemetry from each action becomes a labeled outcome, and when drift shows up, the model retrains on it automatically. Entities, actions and roles: the same model behind every job.

Closed loop it only gets better SIMULATE ACT ALERT Test strategies on the twin Deploy the winning Action Vector Monitor live · drift · churn

Three twins, one vocabulary your agent already speaks.

Every Cortex forecast is built from three kinds of twin: Entity, Action, and Role. It's the same model behind every job your agent does: the things it acts on, the changes it runs, and the goal it's accountable for.

01
Entity Twin

The blueprint, what your agent acts on

Your client's customers, SKUs, accounts, or assets become queryable twins built from their real data. Your agent can segment, classify, and predict against them, the risk-free sandbox for every what-if it wants to run.

baseline locked · ready to simulate
02
Action Twin

The change, what your agent runs

A task, offer, message, or process change the agent ships becomes a testable action. Replay it across the entity population in simulation, compare scenarios, then push the winning one live when the numbers hold.

executing · live in production
03
Role Twin

The pulse, the agent itself

The agent is a Role Twin: it owns a goal, watches every entity and action, alerts on drift, recommends the next move, and feeds real outcomes back so the next forecast is sharper than the last.

drift detected → recalibrating simulation

From cold start to live forecaster: three stages.

A Cortex model moves through three stages over its life. You don't need labeled data to start exploring, but you do need roughly 200–500 labeled outcomes to leave cold start and ship calibrated predictions. After that, the model never really stops learning.

01
TRAIN

Cold start on your client's history

Point Cortex at first-party data and it trains a calibrated model. Cold start needs ~200–500 labeled outcomes; below that, predictions ship wider confidence intervals and Cortex tells you so.

02
SIMULATE

Run what-ifs before anything ships

Your agent calls the simulate endpoint to replay an action across the entity population and read a distribution of outcomes, with confidence intervals, not a guess. The change fails safely here, where it costs nothing.

03
LIVE

Predict, monitor, retrain on drift

In production the model serves predictions, watches the outcomes that come back as telemetry, and retrains automatically when drift crosses threshold. Every labeled outcome makes the next forecast better.


xG for your AI agents.

In football, expected goals (xG) scores chance quality, not just shots taken, a tap-in and a half-court heave both count as one shot, but only one was ever going in. Cortex does the same for your agent's work. Completion is the shot count: the agent finished the task. Expected Outcome is the chance quality: the probability the action actually worked, normalized by task type so a hard task and an easy one are scored on a fair scale. No failure signal inside the labeling window counts as a success.

Completion

  • Counts that the agent finished the task
  • Every task scored the same, a tap-in equals a long shot
  • Says nothing about whether the action worked
  • No sense of difficulty or context
  • Looks great right up until the outcome lands

Expected Outcome

  • Scores the probability the action actually worked
  • Normalized by task type, hard and easy on a fair scale
  • Measured against the real outcome in the labeling window
  • No failure signal in the window = success
  • Calibrated, with confidence intervals, and it compounds

Completion tells you the agent took the shot.
Expected Outcome tells you whether it was ever going in.


Not another agent.
A forecaster, inside yours.

Get an API key, point your agent at the endpoint, and ship demand, revenue and outcome forecasts, with what-if simulation and drift monitoring built in, under your brand.

Want to talk it through first? Email us at [email protected].

LOCAL-FIRST SELF-HOSTABLE REPRODUCIBLE