TALK TO SALES
TRAINING

Define the target. Agents keep training toward it.

Define the target and the benchmark that measures it. Agents run the training on dedicated GPUs, read every result and keep improving the recipe, round after round.

One Auto-RL job: 6 agents train Qwen3.5-9B with reinforcement learning, one training run each a round.

ONE JOB
RUN BY:
AGENTS
GPUS:
DEDICATED
DATA:
YOURS

How a training job runs

  1. YOU DEFINE THE TARGET

    What the model must get better at, and the benchmark that measures it.

  2. AGENTS WRITE THE RECIPE

    Each round, each agent sets the recipe for one training run.

  3. EVERY RUN IS SCORED

    Each run trains on GPUs and is scored on prompts it did not train on. The agents read the scores and change the recipe.

  4. THE BEST CHECKPOINT IS KEPT

    A checkpoint that fails a check is rejected. You get the best one that passed every check.

A1A2AGENTS A1, A2WRITE EACH RECIPETRAINING RUNSSCORED EVERY ROUNDTHE TARGETYOUR BENCHMARKKEPT CHECKPOINTPASSED EVERY CHECK
METHODS

Every method, the same agents and tools

  1. POLICYAGENTS RUN ITROLLOUTSUPDATEREWARD YOU DEFINE
    REINFORCEMENT LEARNING

    Reward the outcomes your work needs.

    Agents run rollouts continuously against the environment and reward you define, and every round of rewards updates the policy.

  2. ROUND AFTER ROUNDKEPT ONLY IF BETTERREJECTEDYOUR TEAM STEERS
    POST-TRAINING

    Agents keep a round only if it beats the last checkpoint.

    Agents start a new round of post-training as soon as the last one is scored, so the model keeps improving after the first job ends.

  3. YOUR EXAMPLESBASE MODELOPEN WEIGHTFINE-TUNINGAGENTS RUN ITTASK OUTPUT
    FINE-TUNING

    Adapt an open-weight model to your own data.

    Agents fine-tune an open-weight model on your examples, score each run on your benchmark, and keep optimizing between runs.

ONE REINFORCEMENT LEARNING JOB
METHOD:
LORA GRPO
STEPS A RUN:
40

Agents took Qwen3.5-9B from 65.8 to 77.0 on JevBench Capability.

JevBench is a set of 231 public questions. Capability is the average of its two scores, Intelligence and Calibration. The agents never see the questions. Decision-1 is a 9B decision model from Microsoft.

+6.2
POINTS ABOVE MICROSOFT DECISION-1, WHICH SCORES 70.8
180
TRAINING RUNS BY ROUND 34, FROM 6 AGENTS
79.0%
BEST RUN ON THE JOB'S EVAL. THE STARTING MODEL SCORED 71.6%
HOW BLACKBOX USES DECISION-1
NEXT · INFERENCE

No one can read your prompts. Not even us.