Hue launches at YC with stateful fake copies of Gmail, Slack and Salesforce for agent testing

An AI agent that writes to a user's inbox or CRM has nowhere safe to test the next change. The live account is off limits. A mocked API does not act like the real app. A fresh sandbox is empty. There are no customers in it. There is no history behind the next action. Hue launched on Y Combinator Fall 2026 today to fill that gap.

The product is a stateful simulation layer, the launch page says. The target apps named on day one are Gmail, Slack, Salesforce, Notion and QuickBooks. Hue says it creates stateful clones of those apps. A team can test changes against realistic tools, data and state before they reach users, the page says. Hue had 34 votes on YC in the first hours of the launch.

Two ways to fill the copy

The launch post names two methods. The first is to generate a world. A coding agent describes the environment to test against. The example in the post reads "the apps of a 40-person software company that just raised their Series B". Hue then fills the connected apps with months of realistic history, within a couple of minutes, the post says.

The second method starts from real runs. Hue uses an agent's traces to recreate the state of the world during a real run, the post says. A team that had a production failure can rebuild the exact state the agent was in. The team can then try a fix against that state as many times as needed. The result can be checked before release.

Who built it

The founders are Aneesh Kethini and Josh Le. Aneesh was the product manager for Datadog's AI SRE agent. Before that he was a quantitative trader at SIG. Josh built frontier legal agents at August. He was a software engineer at Apple, Gusto, SendBlue and Wayfair before that. The launch page says both have years of work on agent testing environments. Both have built systems that turn production usage into replayable runs.

The company's own one-line summary names the goal. Hue creates stateful clones of the apps the agent uses, the launch page says. The industries list on the page is Developer Tools, SaaS/Subscription Software, Enterprise Software and B2B.

What a buyer tests

A reader with an agent in production has a clear test to run. Pick one real failure that was hard to reproduce. Replay the trace into Hue. Fix the agent. Rerun from the same starting point. Measure.

Then do the other direction. Describe the kind of customer company that most of your agent's traffic comes from. Have Hue fill the apps with a world that fits that description. Run the agent against that world and watch where it misbehaves. The founders ran that work at Datadog and August, the launch page says.

The limit is on the fidelity of the clone. Gmail is one well-known target with public behaviours. A Salesforce tenant has custom objects, workflows and permission rules that differ for every customer. A clone that lacks a given field or rule will miss a class of failure that only shows in production. The buyer has to measure how close the clone sits to the real customer tenant before trusting the test result.

Source

Hue: Simulations for AI Agents, Y Combinator Launch YC post.