Meet Ultraresearch — an autonomous AI research system designed for frontier-scale autoresearch. Dynamically coordinate a diverse pool of continuously self-improving agents, flowing toward north stars on the horizon of superintelligence.
Ultraresearch
Ultraresearch is Ultraresearch’s autoresearch infrastructure for the age of autoresearch. It is designed to conduct complex research autonomously on problems that cannot be solved through scaling existing patterns. Most current systems remain constrained by patterns present in their training data and lose effectiveness once a problem moves beyond that distribution. Given an objective, Ultraresearch first converts the goal into a formal specification through autoformalization. It then runs kernel-driven exploration by coordinating a small number of specialized approach agents. These agents attack the problem from different angles, generate and test candidates at scale, and continuously synthesize the resulting evidence. The system is continuously self-improving.
As new evidence is produced and verified, it refines its strategies and internal processes over time. Ultraresearch operates end-to-end and maintains full transparency: it can articulate its current approach, explain the reasoning behind its decisions, track progress, and clearly present verified results. Its purpose is to automate frontier research through structured orchestration and rigorous verification, enabling small teams to tackle problems that would traditionally require significantly larger organizations.
Every run driven by the Three-Timescale Learner.
Ultraresearch doesn’t replay what it was trained on—it keeps learning. Point it at a hard objective and it improves across the run, carrying forward what it discovers instead of starting cold every time.
- 01
Beyond the corpus
Most systems stall the moment a problem outgrows their training data. Ultraresearch is built to keep reasoning past the edge of what it has already seen.
- 02
Sample-efficient
It picks up a new domain from a handful of examples rather than a purpose-built dataset—closer to how a person learns a new skill.
- 03
Self-directed progress
A robust internal sense of what is working keeps long runs on track, without constant supervision or hand-holding.
- 04
Compounding skill
What one run learns is retained and shared across others, so capability accumulates instead of resetting to zero.
- 05
Improves on the job
Deployment is not a frozen snapshot. Ultraresearch keeps getting better as it works, closing its own gaps instead of waiting for the next release.
- 06
Verified, not hopeful
It charts its own path toward your objective and returns only what survives rigorous verification.
Recent Research
View all- We’re live! First research notes are cooking, stay tuned
The objective is yours.
What will you build next?
- Open Source
- GitHub