Keep your harness.Own the model that drives it.
Your agent improves itself from its own work, round after round: better, faster and cheaper the more it runs, in a model you own.
Better, faster and cheaper, the more it runs.
2.7×
the tasks solved
½
the steps it takes per task
96%
of what lasts, kept while it learns what changes
One base URL.
Point your harness’s model calls at unify and add one header. Your prompts, tools and provider key stay as they are, and every call becomes something your model learns from.
from openai import OpenAI
client = OpenAI(
base_url="https://api.unify.ai/openai/v1",
default_headers={"X-Unify-Key": UNIFY_KEY},
)Recursive self-improvement, for the agent you run.
- CaptureYour agent’s history first, then its live calls, through one base URL.
- ImproveYour users’ feedback, hindsight and frontier models judging every step show your model what to do better.
- ReleaseNew weights take your calls once they’re within the margin you set. Hold or roll back at any time.
The harness becomes its training pipeline.
Today an agent learns in its harness, as notes, skills and memory files that only grow. Here, what it learns moves into the weights, round after round, and the harness stays thin.
No setupno reward to define, no labels to write, no environment to build
Your weightsserved for you, or yours to run anywhere, on-premises included
Your networkself-host the gateway and the training, and your calls stay inside it
Your first round trains on your history.
We set it up with the engineer who owns your harness: how your past and live calls reach us, and the gain you’re after. Your users meet the new model only once it’s within the margin you set.