Skip to main content

WootzApp / Agent traces. Work traces.

Ground truth
others can't trace.

Agent traces and work traces, captured below the harness and inside the browser.

An OpenShell-based hypervisor records execution below the harness. W8, our own Chromium browser, records the page state screenshots miss. Our team turns those traces into datasets, evals, and RL environments for frontier model labs, and custom small models for enterprises.

Bring your production harness. Change almost nothing.

Data and environments for labsCustom models for enterprises

Our data passes the quality gates of the world's leading frontier model labs.

Read the published benchmark
runtime / event streamIllustrative event format
model.request     route=gateway  status=200
proxy.decision    dest=api.example.com  ALLOW
workspace.write   /sandbox/work/src/app.ts
egress.attempt    dest=203.0.113.9:443  DENY
policy.event      binary=/usr/bin/curl  blocked
Evidence scope is configured per sandbox policy.
browser / dom_model10.txtRecorded benchmark excerpt · Task 2
  • Developed by → Tiburon
  • Release date → 8/13/2026

Both values were absent from the screenshot record of the same run.

View the Task 2 evidence audit ↗

What we build / Two products, two observation points

Training data for frontier labs. Custom small models for enterprises.

  • HypervisorOpenShell-based execution layer. Observes the agent from below the harness.
  • W8 browserOur own Chromium browser. Records the page state screenshots miss.

For frontier model labs

Datasets, evals, and environments from traces nobody else captures.

We record work traces from experts and agent traces from models inside our hypervisor and the W8 browser. Our team turns them into datasets, evals, and resettable environments that beat state-of-the-art models. Rewards read page state and the runtime record directly, so they score what the agent did, including the attempts the sandbox stopped.

What the capture layer adds

  • W8 browserExact values, control states, and off-screen content at every step, in a compact DOM-diff record.
  • W8 browser92.3% fewer evidence gaps than screenshot recordings, across 100 published tasks.
  • HypervisorReward-hacking evidence. Denied egress, policy decisions, and workspace file changes are recorded below the agent, whether or not the agent logs them.
  • HypervisorModel calls captured at the gateway, in one view with runtime and browser events.
  • HypervisorW8 browserDeterministic checks wherever the outcome is observable. Model judges cover visual and subjective criteria.

Delivered as: Trace datasets, evals, and resettable environments with fixtures, task specifications, verifiers, and reward functions.

Our data passes the quality gates of the world's leading frontier model labs.

Discuss data and environments

For enterprise AI teams

Small models trained on what your agents actually do.

Your harness runs unchanged inside our hypervisor and produces agent traces from your real tasks. W8 verifies the browser outcomes. The team that builds data for frontier labs trains small models on the traces that verifiably succeeded.

What the capture layer adds

  • HypervisorNo harness rewrite and no tracing SDK. Package it, configure the sandbox and gateway, and run it.
  • HypervisorTraining data from below the harness: processes, egress, policy decisions, and workspace changes alongside every model call.
  • W8 browserBrowser tasks verified against DOM state, so the training set holds the runs that completed.
  • HypervisorW8 browserRewards built into the capture layer: deterministic checks on browser and runtime state.
  • HypervisorRouting between your small model and frontier models by verified task success, at the gateway where requests originate.

Delivered as: Fine-tuned small models, evaluations built from your real tasks, and routing at your gateway.

Your agent doesn't need to change. Your visibility does.

“Almost nothing” means core agent logic stays as written. Setup covers image packaging, runtime configuration, credentials, and gateway routing.

Discuss a custom model

How the hypervisor and W8 work →

Published research / W8 DOM-diff Bench · Oct 2026

92.3% fewer recorded evidence gaps.

We ran two verification pipelines on the same 100 browser-agent tasks: one reading screenshots; the other reading structured browser evidence captured by W8. We held task inputs, action histories, final answers, and scoring rubrics fixed.

100 tasks. 639 criteria. 92.3% fewer paired source-evidence gaps in the W8 recordings.

Read the full study and per-task resultsSource and reproduction
Recorded representationCriteria with lossRate
Screenshots26 of 6394.07%
W8 DOM-diff Optimized2 of 6390.31%
Screenshots4.07%
W8 DOM-diff Optimized0.31%
Criteria with asymmetric source-evidence loss, shared 0–5% scale. Lower is better.

The metric counts criteria whose evidence is present in one recorded format and missing from the other. Agent success, verifier accuracy, and training lift are separate measurements. This comparison uses LLM-based grading; objective browser-state rewards are the subject of a planned benchmark.

From trace to training data

Captured by our infrastructure. Engineered by our team.

Agent traces come from AI agents running your tasks. Work traces come from experts doing the same work. Both are recorded below the harness and inside the browser, where nobody else captures them.

Our team and infrastructure turn these traces into datasets, evals, and environments that beat state-of-the-art models.

  • Hypervisor

    Captured below the harness

    The hypervisor records processes, egress, policy decisions, and workspace file changes from outside the agent. What the agent logs is one input among several.

  • W8 browser

    Recorded inside the page

    W8 stores page state at every step as a DOM-diff, with exact values, control states, and off-screen content.

  • WootzApp team

    Engineered by our team

    Our researchers and engineers turn raw traces into datasets, evals, and resettable environments, with task specifications, verifiers, and rewards built on browser and runtime state.

  • WootzApp team

    Through frontier quality gates

    Every delivery is checked against explicit task criteria. Our data passes the quality gates of the world's leading frontier model labs.

01 / The Agent Hypervisor

The harness reports what it knows. The runtime reveals what it does.

An agent can call tools, spawn processes, read files, attempt connections, trigger policy denials, and recover from failures without those events being faithfully reflected in its model transcript.

WootzApp's execution layer builds on NVIDIA OpenShell and the Openrind Shell runtime. It keeps existing agent logic intact and brings model interactions, configured runtime evidence, and sandbox-policy events into one view.

  • Keep the harness. Preserve prompts, tools, orchestration, and control flow; configure its sandbox, policies, and compatible model gateway.
  • Observe outside its self-report. Correlate available process/workspace and network-policy evidence with model and tool events.
  • Capture security-relevant attempts. Record denied egress and policy decisions alongside the tool calls the agent chose to log.
  • Improve where requests originate. The controlled inference path is the integration point for model selection, tracing, and operator-approved routing changes.

Keep your agent's framework. Move the observation boundary.

Explore the OpenShell-based architecture

OpenShell sandbox

Your existing harness

Prompts · tools · orchestration · control flow

unchanged agent logic

Observation boundary

supervisor · Landlock + seccomp · network namespace · local proxy

Evidence stream →

  • Model requestGateway route
  • Proxy decisionBinary-attributed endpoint policy
  • Workspace updateFUSE mount · /sandbox/work
  • Egress attemptSandbox policy · allow / deny
Captured events follow the sandbox configuration and its instrumentation. Built on NVIDIA OpenShell; no NVIDIA endorsement implied.

02 / W8 Browser-Native Evidence

A screenshot records the frame. We record the page.

A screenshot can miss a value outside the viewport, lose the relationship between a label and a field, or fail to preserve the state needed for a rubric check.

W8 is built on custom Chromium infrastructure. Its DOM-diff Optimized recording preserves compact structured browser states: content, controls, relationships, and scroll context, alongside screenshots and executed actions.

  • Read what pixels miss. Preserve exact values, control states, and additional page content when available to the browser capture.
  • Connect an action to a state. Inspect ordered page states before and after each interaction, independent of the agent's final answer.
  • Make verification inspectable. Evaluate recorded evidence against task-specific criteria, with deterministic state checks where ground truth can be expressed objectively.
  • Prepare stronger learning signals. Retain the evidence needed to audit trajectories, failures, and candidate RL rewards.

Richer evidence for the judge. Explicit state checks where a criterion allows them.

Inspect W8 DOM-diff
Frame · screenshot10.pngAudit summary
Gallery
Description
viewport ends

developer · release date — not in frame

Page · dom_model10.txtRecorded excerpt
  • Developed by → Tiburon
  • Release date → 8/13/2026
Task 2, Xbox cross-site research. Evidence audit ↗ DOM-diff Optimized is a compact per-state representation. W8 keeps screenshots for visual, canvas, and geometry-dependent criteria.

A model endpoint sees what the model returns.

A harness trace sees what the harness emits.

A screenshot sees what the camera captured.

Others train on what your harness exposes. We reveal what your harness cannot see.

That's the evidence other layers cannot independently establish.

Deterministic rewards

Deterministic rewards. First-class infrastructure.

A trace records what an agent attempted. A verifier decides whether it worked. We build the verifier into the capture layer: browser-state checks from W8, runtime checks from the hypervisor. Visual and subjective criteria keep model or human grading.

  1. 01HypervisorRunRun the existing agent or an authorized human workflow inside the configured environment.
  2. 02HypervisorW8 browserObserveObserve supported runtime events, model/tool activity, and W8 browser states.
  3. 03W8 browserVerifyVerify outcome evidence against expert-defined rubrics or machine-checkable state assertions where applicable.
  4. 04HypervisorW8 browserTrainOur team builds datasets, evals, and environments for model labs, or trains and evaluates a small model for an enterprise.

The same capture layer serves human experts demonstrating work and agents attempting work. The instrumented execution substrate stays constant across both.

Inputs
  • Existing agent
  • Expert operator
Observation points
  • OpenShell execution layer
  • W8 browser
Evidence + verifiers
  • Runtime events
  • Browser states
  • Rubrics · state checks
Products
  • Datasets, evals, environments · model labs
  • Fine-tuned small models · enterprises
We turn a captured trajectory into a resettable RL environment by adding explicit fixtures, reproducible resets, and validated reward implementations.

Research / Published and planned

Show the missing evidence.

Our first published study measures where screenshots lose criterion-relevant browser evidence, using paired trajectories and a frozen evaluation protocol.

The planned studies ask three questions. How much execution evidence does the hypervisor preserve beyond SDK traces? How accurately can objective browser-state checks replace subjective screenshot judgments? Does that improve model training per dollar?

Each answer ships with artifacts and a controlled benchmark.

Explore WootzApp research
  1. 01Published · Oct 2026

    Browser evidence fidelity: screenshots vs W8 DOM-diff Optimized

    100 tasks, 639 criteria. Screenshots lacked criterion-relevant source evidence on 4.07% of audited criteria; W8 records on 0.31%.

    Code ↗Dataset ↗Per-task results ↗Limitations ↗
  2. 02Planned benchmark

    Hypervisor capture completeness vs SDK traces

    How much execution evidence does the OpenShell-based layer preserve beyond what SDK traces report?

  3. 03Planned benchmark

    State-based rewards vs LLM judges

    How accurately can objective browser-state checks replace subjective screenshot judgments?

FAQ

Harnesses, judges, security, and data.

What does WootzApp sell?

Two products built from agent traces and work traces. Datasets, evals, and RL environments for frontier model labs, engineered by our team from traces recorded inside our hypervisor and the W8 browser. Our data passes the quality gates of the world's leading frontier model labs. Custom small models for enterprises, trained on traces of your own agents that the hypervisor recorded from below the harness and W8 verified in the browser.

What is a work trace?

A recording of an authorized expert doing a real task inside our hypervisor and W8: their actions, the runtime events, the page state at every step, and the assessed outcome. An agent trace is the same record produced by an AI agent.

Do we need to change our agent?

No rewrite of core agent logic and no tracing SDK. Setup is image packaging, sandbox configuration, and gateway routing. From there the hypervisor observes the agent from below the harness, independently of what the agent logs, and W8 records the browser state its tasks depend on.

Does W8 remove the need for LLM-as-a-judge?

For some criteria. Criteria that reduce to stable, observable application state get explicit checks. Visual and semantic criteria keep model or human grading. The published benchmark measures evidence preservation and uses model-based verifiers.

Is this an agent security product?

Security is one application. The product is execution-native ground truth for training and evaluation, and security-sensitive runtime attempts feed red-team tasks and training. Policy enforcement and runtime-event collection follow the configured OpenShell integration.

Do you resell enterprise data?

Not by default. Enterprise-private training and licensed model-lab data are separate contracts and collection programs. External licensing requires explicit rights, authorized environments, and permitted use.

What your agent did matters more than what it said.

Bring the harness. Bring the hard tasks. We'll show you the missing evidence.

Talk to engineeringRead the research