Hue
Hue
Hue
Every production run, ready to run again.
Every production run, ready to run again.
Every production run, ready to run again.
Backed by investors from
Backed by investors from
Backed by investors from
A production run happens once. The tools answer, the data moves, and the moment is gone. Hue keeps it.
A production run happens once. The tools answer, the data moves, and the moment is gone. Hue keeps it.
A production run happens once. The tools answer, the data moves, and the moment is gone. Hue keeps it.
01
01
01
Fidelity
Fidelity
Fidelity
Everything the agent saw, everything it touched, and the ability to do it again.
CONTEXT
01
TOOLS
02
OUTCOME
03
RERUN
04
COMPARE
05
KEEP
06
WORLD
07
Everything the agent saw, everything it touched, and the ability to do it again.
CONTEXT
01
TOOLS
02
OUTCOME
03
RERUN
04
RISK ASSESSMENT
05
DECISION MAPPING
06
VENDOR BAKEOFFS
07
Everything the agent saw, everything it touched, and the ability to do it again.
CONTEXT
01
TOOLS
02
OUTCOME
03
RERUN
04
COMPARE
05
KEEP
06
WORLD
07
Hold the world constant. Change the agent.
Hold the world constant. Change the agent.
Latency
Accuracy
Cost
Divergence
When you change the agent, where does the rerun split from the original: which turn, which tool call, which decision?
Distribution
Divergence
What does the work actually look like, and how much does it vary from one case to the next?
When you change the agent, where does the rerun split from the original — which turn, which tool call, which decision?
Reliability
Accuracy
How often will the agent get it wrong — and what are the stakes when it does?
How often does each model or harness version get the real scenario right, and which runs flip when you change something?
Cost
What does the work cost to run, and how does that curve look plotted against accuracy?
What does each run cost across models and versions, and where does spend actually buy accuracy?
What does each run cost across models and versions, and where does spend actually buy accuracy?
Latency
How fast do answers come back once the work is running at real volume?
How long does each run take end to end, tool calls included, for every variant you try?
02
02
02
Value
Value
Value
Every change to your agent can be tried against the exact runs your customers already had, under the exact conditions they had them.
Reproduce any failure
A customer says the agent got it wrong. Pull the run, watch it fail the same way, fix it, run it again. No guessing at what the tools returned.
Reproduce any failure
A customer says the agent got it wrong. Pull the run, watch it fail the same way, fix it, run it again. No guessing at what the tools returned.
Reproduce any failure
A customer says the agent got it wrong. Pull the run, watch it fail the same way, fix it, run it again. No guessing at what the tools returned.
Per-run truth
Not one aggregate score. For each real run: did it succeed, what did it cost, how long did it take, where did it diverge.
Per-run truth
Not one aggregate score. For each real run: did it succeed, what did it cost, how long did it take, where did it diverge.
Per-run truth
Not one aggregate score. For each real run: did it succeed, what did it cost, how long did it take, where did it diverge.
Swap anything
Try a new model, a new prompt, a new tool schema against the same preserved run. The difference you see is the change you made.
Swap anything
Try a new model, a new prompt, a new tool schema against the same preserved run. The difference you see is the change you made.
Swap anything
Run competing models and harnesses against the same environment and compare outcomes on the work that matters.
Fidelity that doesn't decay
Tools get deprecated, data moves, APIs change. A preserved run doesn't. A run from March reruns the same way in December.
Fidelity that doesn't decay
Tools get deprecated, data moves, APIs change. A preserved run doesn't. A run from March reruns the same way in December.
Fidelity that doesn't decay
Tools get deprecated, data moves, APIs change. A preserved run doesn't. A run from March reruns the same way in December.
Nothing approximated. Nothing mocked by hand. The run, as it happened, ready to run again.
Tool responses preserved, byte for byte
Live API calls needed to rerun a scenario
Thing that changes between reruns: your agent
03
03
03
Research
Research
Research
Notes on why benchmarks drift from reality, why agent runs are hard to reproduce, and what fidelity actually requires.
OpenAI
February 23, 2026
The industry’s coding benchmark is contaminated
OpenAI
February 23, 2026
The industry’s coding benchmark is contaminated
METR
July 10, 2025
Felt 20% faster, measured 19% slower
METR
July 10, 2025
Felt 20% faster, measured 19% slower
Gartner
June 25, 2025
40% of agentic AI projects will die
Gartner
June 25, 2025
40% of agentic AI projects will die
Liang, Garg & Moghaddam
June 14, 2025
Models memorized the benchmark, not the skill
Liang, Garg & Moghaddam
June 14, 2025
Models memorized the benchmark, not the skill
04
FAQ
Answers to the questions agent builders ask before they trust a rerun.
01
What does Hue actually preserve from a run?
01
Everything the agent saw and everything it touched: the prompts and context in order, every tool call and its exact response, and the outcome. Enough to rerun the scenario with a different agent and get a fair comparison.
02
My tooling already runs evals on production data. What's different?
02
Those tools re-score a run: they grade what already happened. Hue re-executes it. The agent runs again inside the same conditions, so you learn how a new model or prompt would have behaved, not how the old run would grade.
03
Does an agent need to be deployed first?
03
Yes. Hue works from real runs, so it needs production or staging traffic.
04
How are third-party tools simulated?
04
Every tool call in the original run is served back with the same response and behavior, so reruns never hit a live service and never drift from the original scenario.
05
What can I compare?
05
Any version of your harness — prompts, tools, orchestration — and any model, against the same preserved run. Accuracy, cost, latency, side by side.
04
FAQ
Answers to the questions agent builders ask before they trust a rerun.
01
What does Hue actually preserve from a run?
01
Everything the agent saw and everything it touched: the prompts and context in order, every tool call and its exact response, and the outcome. Enough to rerun the scenario with a different agent and get a fair comparison.
02
My tooling already runs evals on production data. What's different?
02
Those tools re-score a run: they grade what already happened. Hue re-executes it. The agent runs again inside the same conditions, so you learn how a new model or prompt would have behaved, not how the old run would grade.
03
Does an agent need to be deployed first?
03
Yes. Hue works from real runs, so it needs production or staging traffic.
04
How are third-party tools simulated?
04
Every tool call in the original run is served back with the same response and behavior, so reruns never hit a live service and never drift from the original scenario.
05
What can I compare?
05
Any version of your harness — prompts, tools, orchestration — and any model, against the same preserved run. Accuracy, cost, latency, side by side.
04
FAQ
Answers to the questions agent builders ask before they trust a rerun.