Tokenless GTM brain

Tokenless helps AI labs and agent evaluation teams find and fix reward hacking and grader flaws in RL environments.

category

RL environment and evaluation auditing

market profile

  • text: Longer agent tasks expose opportunities to exploit graders or environment design.
  • text: A flawed reward can make apparent model progress misleading and weaken confidence in training.

personas

  • title: RL post-training lead
    pains:
    • Optimization amplifies exploitable reward flaws
    • Cannot distinguish real capability gains from gaming the grader
  • title: Evaluation infrastructure engineer
    pains:
    • Complex agent tasks introduce hidden grading bugs
    • Manual audit of environments takes scarce engineering time
  • title: AI safety research lead
    pains:
    • Unintended agent behavior can undermine alignment experiments
    • Weak evaluation validity obscures deployment decisions

icp

  • text: Frontier labs with RL post-training and internally developed training environments.
  • text: Agent evaluation builders maintaining complex tasks and automated graders.

message pillars

  • text: Find and fix evaluation weaknesses.
  • text: Prevent model exploitation of flawed environments.
  • text: Increase confidence in model results.

objections

  • objection: Our evaluation engineers already audit graders.
    response: Compare findings on a scoped environment with the internal review and reproduce each issue.
  • objection: Will an outside auditor handle proprietary environments safely?
    response: Agree access controls, data handling and a bounded test before sharing sensitive tasks.

competitors

  • name: Gray Swan AI
    why: Adjacent safety evaluator offering adversarial evaluation and reward-hacking benchmarks.
  • name: Vals AI
    why: Independent evaluation provider competing for model and agent evaluation budgets.
  • name: METR
    website: https://metr.org/blog/2026-05-19-frontier-risk-report/
    why: Independent evaluation research organization with environment hardening expertise.

influencers

  • name: Kevin Wang
    url: https://nitter.cf/kevzwang
    why: Founder discussing the company's evaluation research.
  • name: Rohit Agarwal
    url: https://teapawt.net/Rohit_Writes
    why: Founder sharing benchmark auditing research.

keywords

  • reward hacking detection
  • RL environment audit
  • AI agent grader bugs
  • benchmark auditing
  • LLM eval integrity
  • RL verifier exploits

voice

  • research-led
  • precise
  • mission-driven

signals

  • text: A lab reports anomalies in RL reward or benchmark results.
  • text: A team is scaling long-horizon agent environments and automated grading.

pricing

model: unknown

motion

hybrid

channels

  • research blog
  • open source
  • X
  • YC directory

confidence

medium

Founders

  • name: Rohit Agarwal
    title: Founder
    role: Founder
    bio: Founder at Tokenless (YC S26). CS PhD student at Princeton PLI, Hazan Lab, Google Deepmind. Previously at Duolingo and theory research at Berkeley, was EECS #1 BS grad in 2024.
  • name: Andrew Liu
    title: Founder
    role: Founder
    bio: Founder at Tokenless (YC S26). Formerly a researcher at UC Berkeley working on language documentation tools. BA in Computer Science, Mathematics, and Linguistics from UC Berkeley. Previously TA'd CS61C.
  • name: Kev Wang
    title: Founder
    role: Founder
    bio: Founder at Tokenless. Prev MTS at SkyPilot building multi-cloud GPU infra for top neolabs, Founding Engineer at Eventual from seed to Series A building the data engine for physical AI, EECS at UC Berkeley.

Back to YC companies