Why we’re building Hue

We want people to be able to hand an agent meaningful work and trust it to do that work well. We’re building Hue to help teams earn that confidence and keep expanding what their agents can take on.

What building agents taught us

We came to this problem through our own work. Aneesh was the PM on Datadog’s AI SRE. Josh built legal agents at August. Both of us helped build and maintain infrastructure to turn production usage into rerunnable tests. Comparing models and catching regressions meant recreating the situations that users asked our agents to handle.

For the AI SRE, that meant archiving incidents and their surrounding metrics, logs, and traces. The environment also had to simulate telemetry the original agent had never queried. A new model or updated harness might follow a different lead or inspect different data to end up at a better conclusion. For the legal agent, this involved the same task but with long-horizon legal tasks: enough of the original situation had to be available and realistic to properly test the agent on the work.

Those experiences shaped our conviction that the work agents encounter in production should be the foundation for improving them. Making that possible takes both scenarios drawn from real usage and environments that faithfully recreate the conditions behind them.

Production usage provides the requests that users make, the context they depend on, and the unexpected circumstances that emerge in practice. It gives teams a set of situations tied to the work their users need done and lets them grow it scalably.

The testing environment needs to preserve the behavior that shapes an agent’s decisions. Tool calls must respond realistically and maintain information relationships. If the agent changes something, later calls need to see that change. These details make it possible to test a new approach under the conditions it would encounter in production.

Replay the work. Expand what’s possible.

We’re building Hue around agent replays: taking a scenario from production and running an updated agent in a reconstruction of that environment. The agent can make new decisions, explore another route, and leave a different result. Developers can try changes against the work that gave rise to them.

We know the effort involved in building and maintaining those environments. We want that infrastructure to be available to more teams, so they can spend more of their time improving what their agents can do.

If you’re building agents for work people depend on, we’d like to hear what you’re trying to make possible.