AI that builds
better AI.
Autonomously.

Each day it comes up with an idea, implements it, evaluates it in the environment, and learns from every experiment, all on its own. Or drop in your own idea, and it biases the search toward that.

See the exploration Add your ideas

NOW BUILDING · CUSTOMER-SUPPORT AGENTS

THE RESEARCH LOG · EVERY EXPERIMENT, ON THE RECORDBETWEEN RUNS

EXP 17 · BANKING · REFINERewrite the agent with a dramatically shorter, priority-ordered system prompt, add programmatic step tracking via injected system messages after list-returning tool calls, fix the user-tool give→instruct→wait flow with f…21%

EXP 01 · RETAIL · EXPLOREReplace the minimal seed system prompt with a deeply policy-aware, workflow-structured system prompt that encodes the complete retail domain policy as precise operational procedures (with step-by-step tool-calling workfl…88%

EXP 01 · BANKING · EXPLOREReplace the minimal seed agent with a retrieval-augmented agent that pre-processes the entire 698-document corpus into a persistent embedding index + discoverable-tool catalog on first use, then on each task proactively…21%

EXP 02 · RETAIL · EXPLOREReplace the minimal system prompt with a comprehensive procedure-driven operational playbook that embeds explicit workflow prerequisites, mandatory tool-use rules (especially for calculate and getproductdetails…78%

EXP 02 · BANKING · EXPLOREBuild a workflow-aware banking agent that pre-classifies customer requests into workflow archetypes (credit card application, cash back dispute, transaction dispute, transfer protocol, referral, account management), proa…24%

GRADED ON HELD-OUT TASKSEVERY ROW: ONE EXPERIMENT, GRADED ON TASKS IT NEVER TRAINS ONSEE THE EXPLORATION ↓

01 · THE LOOP

Autonomous AI research loop.

Better AI system, built by an AI researcher that runs the whole loop: ideate, implement, experiment, evaluate, learn, repeat. Every attempt is preserved and every result feeds a growing archive of knowledge that steers the next idea, so the loop runs open-ended toward a measurably better system.

SCROLL TO EXPLORE THE FULL LOOP →

02 · TRACK RECORD

Verified in the open.

EXHIBIT A · MLE-BENCH · OPENAINº1 OPEN SOURCE

Machine-Learning Engineering

Given a new problem, it builds the entire solution on its own, from raw data to a trained, tuned model, and outperforms every other open agent.

MEDAL RATE

Leeroo

50.7%

R&D-Agent

35.1%

AIRA-dojo

31.6%

ML-Master

29.3%

AIDE

17.1%

MEDAL RATE BY TASK COMPLEXITY

LEEROO BEST REPORTED

Low

68.2%

Medium

44.7%

22.0%

High

40.0%

24.4%

BEST REPORTED PER TIER: R&D-AGENT (LOW) · AIRA-DOJO (MEDIUM) · ML-MASTER (HIGH)

HIGHEST MEDAL RATE AMONG OPEN, REPRODUCIBLE ML-ENGINEERING AGENTS

EXHIBIT B · ALE-BENCHELO 1909 · ▲ 30 VS BEST

Long-Horizon Algorithm Design

The problems have no perfect answer. It designs an algorithm, then rewrites it for hours to push the score higher, head to head with human experts.

LEEROO BEST REPORTED

Gravity Sorting

2528

2446

Coverage Design

2310

2880

Path Planning

2194

2153

Logistics

2040

1965

Signal Encoding

2022

1457

Urban Planning

1980

Route Planning

1839

1740

Network Design

1607

1652

Load Balancing

1353

1331

Zoning

1221

1189

100020003000 ELO

RATED AGAINST HUMAN CONTESTANTS

EXHIBIT C · ONE SAMPLE, END TO ENDCLICK A NODE FOR ITS EXPERIMENT

Open-Ended Optimization Problem.

A sample of the loop on the hardest class of problems: open-ended optimization with no known best answer. This one is drawn from a live AtCoder contest. Encode up to a hundred messages as networks; the channel randomly rewires each one and erases its labels; the receiver must still decode it. There is no formula for the best design. The loop invented its own, rated on the contest's human ladder.

VIEW THE PROBLEM ON ATCODER ↗

2022ELO▲ 762 VS AVERAGE HUMAN

AVERAGE HUMAN 1260 · ALE-BENCH