Home›Companies›Hue

Realistic worlds for agent evaluation

Hue turns production usage of agents into repeatable tests that preserve the tools, data, and state behind the real run. As AI agents take on real work, their behavior depends increasingly on the nuances of the world around them. With Hue, teams can test and improve their agents under conditions that mimic real usage.
Active Founders
Aneesh Kethini
Aneesh Kethini
Founder/CEO
Co-founder & CEO of Hue; Second AI PM at Datadog, built and launched Datadog's AI SRE; Prev Quantitative Trader at SIG.
Josh Le
Josh Le
Founder/CTO
Co-founder & CTO at Hue; Prev engineering at Apple, Gusto (W12), August, Wayfair, SendBlue (S23), Sixtyfour (X25); KP fellow
Company Launches
Hue: Simulations for AI agents
See original launch post

Hi everyone, we're Aneesh and Josh.

TL;DR: Hue creates stateful clones of the apps your agent connects to, so you can test changes against realistic tools, data, and state before they reach users.

The problem

Agents take on real work now: they send emails on our behalf, update CRMs, and issue refunds. But there’s usually no safe place to test them. You can’t experiment on a user’s live inbox, mock APIs aren’t representative, and sandbox tenants start empty, so your agent never meets the messy data and business nuance it will see in production. When an agent fails in production, it’s difficult to reproduce the issue and verify a fix.

How Hue solves this

Hue simulates all the apps your agent might use, like Gmail, Slack, Salesforce, Notion, and Quickbooks. The apps are populated in two ways

  • Generate worlds: let your coding agent describe the environment to test against, like the apps of a 40-person software company that just raised their Series B, and Hue populates the apps with months of realistic history within a couple minutes.
  • Start from real runs: Hue uses your agent traces to recreate the state of the world during a real run, so you can reproduce agent failures and verify the fix.

Your agent uses these apps just like it uses the real ones, without touching any live accounts. You can run your agent from the same starting point as many times as you need, test it against various situations in minutes, set up evals to measure performance quantitatively, and compare versions before you ship.

https://youtu.be/tvGNnKpd8oE

The team

We’ve spent the past few years shipping frontier agents.

  • Aneesh Kethini: Prev PM for Datadog’s AI SRE Agent, and Quantitative Trader at SIG
  • Josh Le: Prev built frontier legal agents at August, SWE at Apple, Gusto, SendBlue, Wayfair

At our previous roles, we were responsible for building high-quality testing environments, infrastructure to turn production usage into replayable scenarios, and systems to simulate representative data. We know first-hand how important realistic environments are to agent testing. Now, we’re building Hue so every team can hand its agent real work and trust it to do that work well.

Our Ask

  • Building an agent that works across third-party tools and still using unrepresentative sandbox tenants or mocked APIs? Book a demo.
  • Know a team shipping frontier agents? We’d love an intro: founders@hue.run
Hue
Founded:2026
Batch:Fall 2026
Team Size:2
Status:
Active
Location:San Francisco
Primary Partner:Aaron Epstein