kReative Labs
Back to kReative Labs

Building 2025 to now · Tech Advisor

Agent evaluation platform

Enterprises are deploying AI agents that take real actions, and most of them cannot answer a basic question afterwards: what did it actually do, and was that good? This is infrastructure for answering both.

Agents are being shipped without a way to judge them

A traditional service either returns the right value or it does not, and a test suite settles the question. An agent chooses a path. It decides which tool to call, how to interpret an ambiguous instruction, and when to stop. Two runs of the same prompt can differ, and both can be defensible.

That breaks the usual release process. Teams cannot tell whether a new model version, a changed prompt, or a new tool made the system better or quietly worse, because there is no stable definition of better and no record of what the agent did on the way to its answer.

In a regulated environment the problem is sharper still. If you cannot reconstruct why an agent did something, you cannot approve it for a workflow that touches patients, payments, or filings.

Capture, evaluate, then gate the release

Agent evaluation platform architecture Agent runtimes are instrumented and emit traces of every action, decision, and output. Traces land in a store, which feeds both automated evaluators and a human review queue. Their scores combine into a comparison layer that detects regressions between versions, which in turn gates the release pipeline and surfaces in a dashboard. 1 CAPTURE 2 EVALUATE 3 GATE Customer agents in their own runtime Instrumentation SDK / middleware Trace store actions, decisions, tool calls, outputs Automated evaluators rules, model-graded, golden sets Human review sampled queue, rubric Combined score per version, per task Version compare regression detection Release gate block on regression Dashboard drill into any trace
The load-bearing idea is the middle column. Automated checks are cheap and cover breadth; human review is expensive and covers judgement. Neither alone is enough, so both feed one score that a release can be gated on.

Capture has to be low friction or nobody instruments anything, so the trace format records what the agent did rather than requiring teams to restructure how they built it. Evaluation is deliberately two-track. Gating is where the value actually lands: a score nobody acts on is a report, not infrastructure.

This diagram describes the approach at the level already public in profile.json and on the portfolio. It is not the internal design document, and implementation detail is deliberately absent while the platform is pre-launch. Shridhar should review and extend this before it is treated as a technical description.

Why this is not just a feature of an observability tool

The obvious objection, and the one worth taking seriously, is that this becomes a checkbox inside an existing LLM observability or MLOps platform within a couple of years. Tracing on its own probably does.

The part that resists absorption is the evaluation layer: rubrics that encode what a specific enterprise means by a good outcome, a human review workflow that survives audit, and the comparison logic that turns all of it into a release decision. That is closer to a quality system than a dashboard, and quality systems tend to be bought deliberately rather than inherited from whichever vendor already had the traces.

The bet is not settled. The agent-facing home page publishes a prompt that asks a visitor's own model to argue the opposite case.

Where it actually is

Stage MVP in progress.
Validation More than 50 enterprise discovery conversations.
Design partners Three pilots underway.
Pricing Seed-stage pricing validated within six months, meaning prospects named a number they said they would pay. Nobody has paid it yet.
Revenue None. This is pre-revenue by design at this stage.
Role Tech Advisor, alongside the day job at Moderna.

Open questions