Neurosymbolic recursive self-improvement|#1 on OSWorld-Verified90.2%|

Agents that fly higher and safer for less.

Icarus is a research lab building the first neurosymbolic recursive self‑improvement system. Runtime and cost efficient enterprise agents with correctness guarantees – see it fly

Received. On our way.

Or email hello@icarus-ai.co

Built by Stanford and UIUC CS PhDs in formal methods and machine learning

A neurosymbolic system that specializes to your work, so every run compounds.

LEARN, GENERALIZE & COMPILE

Learned once, compiled into programs.

As it learns how your enterprise operates, Icarus compiles repeated work into primitives (code, smaller task-specific models, subagents). The primitive bank grows and specializes to your roles, systems, and rules, and it generalizes on its own to related tasks and to new teams on the same systems. So every task it learns makes the next one faster, cheaper and more reliable.

RUN 0148 actions RUN 025 primitives RUN 031 program

CONTINUOUS OPTIMIZATION

Recursive self‑improvement, runtime and cost efficient.

As it learns how your enterprise operates, Icarus rewrites its own execution: verified stretches become programs, small specialized models and verified subagents, so the same result stops paying for frontier reasoning twice.

OSWORLD-VERIFIED90.2%#1 verified score COST REDUCTION95%+on repeated tasks

QUALITY AND EFFICIENCYAug 2026

SCORE · OSWORLD-VERIFIED369 real desktop tasks

OSWORLD-VERIFIED90.2%#1 score

Qwen3.8 Max open86.1%

Claude Mythos 5 closed85.0%

Claude Fable 5 closed85.0%

#1 OVERALL+4.1 pts over next best

TASK COST · QUARTERLY COMMISSION CLOSEValidated on real long-horizon enterprise tasks

SAME RESULT, BY VERIFIED RUN

Run 01 reasoned19m 23s$5.29

Run 02 3 parts reused6m 27s$0.80

Run 03 5 parts reused4m 08s$0.07

RESULT PARITY · 100%5 / 5 CHECKS EACH RUN

OSWORLD-VERIFIED#1 SCORE

90.2%#1 SCORE

Qwen3.8 Max 86.1%Claude Mythos 5 85.0%Claude Fable 5 85.0%

REPEATED ENTERPRISE TASKS95%+ COST REDUCTION

RUN 0119m 23s$5.29

RUN 026m 27s$0.80

RUN 034m 08s$0.07

RUNTIME VERIFICATION

Correctness guarantees, checked every step.

From your policies and each task’s stated intent, Icarus synthesizes checks and applies them to every action as it runs: formally where a condition can be stated precisely, by model review where it can’t. Incorrect and unsafe actions are pruned before they execute to maximize efficiency and improve accuracy.

CONDITIONS MET3 / 4run 08 HELD1before write

Quarterly renewal review / Run 08unsafe action held

REQUIRED CONDITIONSlearned from 12 runs and Policy v3.2

✓

Every in-scope renewal is included4 Salesforce records matched to 4 brief entries

met

✓

Usage uses the same 90-day periodQuery and brief both reference Jun 1–Aug 29

met

✓

Support cases remain linked to sourceCase IDs CS-1842 and CS-1911 preserved

met

Discounts above 15% require approvalBeacon Energy requests 18%; held for R. Devi

held

FORMAL 2 · REVIEWED 2CHECKED DURING EXECUTION, NOT AFTER

QUARTERLY RENEWAL REVIEW · RUN 08UNSAFE ACTION HELD

REQUIRED CONDITIONSCHECKED DURING EXECUTION

✓Every in-scope renewal is included

✓Usage uses the same 90-day period

✓Support cases remain linked to source

!Discounts above 15% require approval18% discount blocked before write

News.

COMPUTER USE2026

Daedalus plays real-time Tetris

The piece falls whether or not the agent has finished thinking, which is why no other computer-use agent can play it.

View project →

1ST PLACEAI4Math · ICML 2026

Won the AI4Math Lean theorem-proving competition

Lean checks every proof mechanically, so the placing is verified rather than judged.

View leaderboard →

Built by top researchers in formal methods and machine learning.

Calvin Xu

Co-Founder · CEO

PhD CS @ UIUC. ML × Formal Methods. Ex-MIT Lincoln Lab, Intern @ Meta/TikTok.

Web ↗ LinkedIn ↗

Debangshu Banerjee

Co-Founder

PhD CS @ UIUC. ML × Formal Methods. Ex-Google, Intern @ Amazon/Google.

Web ↗ LinkedIn ↗

Tarun Suresh

Co-Founder

PhD CS @ Stanford. ML × Formal Methods. Ex-Intern @ Nomic AI/Bloomberg.

Web ↗ LinkedIn ↗

Gagandeep Singh

Advisor

Asst. Prof. UIUC. Sponsored by NSF, Amazon, Google, Anthropic, Qualcomm and Bloomberg.

Web ↗ LinkedIn ↗

§ Join

Put verified agents on work that has to be right.

We’re working with a small number of enterprise design partners where agents act across applications, data and approvals, and the work has to be both efficient and provably correct.

Received. On our way.

Or email hello@icarus-ai.co directly.