Agent traces · Work traces
Ground truth others can't trace.
A model proxy sees none of this.
Our team turns both records into training data for frontier labs and small models for enterprises.
What screenshots lose against W8 page state, measured on the same 100 agent runs and 642 audited rubric criteria.
Browser tasks on the public web, each with a screenshot run and a W8 run, plus the raw recordings and saved evals behind W8 DOM-diff Bench.
Post-training a coding model with execution as the grader. Compiler errors and failing tests build the repair practice. Executable checks score every answer.
Runs coding agents in an OpenShell sandbox. Project files live on a FUSE mount backed by PostgreSQL.
For research and data partnerships: founders@wootzapp.com
Where the data comes from
Other training platforms record at the model proxy. They keep what the model was asked and what it answered. Our hypervisor and W8 browser record what the agent did and what the page held. A proxy trace holds three of the ten kinds of evidence below. Ours holds all ten.
We built both recorders. How they work
| What the traces hold | Othersmodel proxy | WootzApp |
|---|---|---|
| Model layer | via gateway | |
| Prompts in | ||
| Completions out | ||
| Tool calls and results the model saw | ||
| Below the harness | hypervisor | |
| Processes | ||
| Network calls, including denied ones | ||
| Each allow or deny | ||
| File changes | ||
| Inside the page | W8 browser | |
| The whole page at each step | ||
| Exact values and control states | ||
| Content off screen | ||
| reaches a proxy trace only if your harness reports it. | ||
For frontier model labs
Recorded below the harness and inside the browser, past where a model proxy stops. They pass frontier-lab quality gates.
For enterprise AI teams
Bring your production harness. Change almost nothing.
How it works
The hypervisor and W8 record each run below the harness and inside the page.
HypervisorW8 browserOur team builds datasets, evals, and environments.
WootzApp teamTo frontier labs, or as a small model for your team.
A model proxy sees what the model said.
We see what the agent did.
Below the harness. Inside the browser.
The technology
An OpenShell sandbox records each process, network call, allow or deny, and file change. It does this whatever the agent logs.
A proxy sees the model call. The hypervisor sees what the call set in motion.
Architecture
Our Chromium browser records what screenshots miss: exact values, control states, and content off screen.
A proxy sees the screenshot the agent sent to the model. W8 sees the whole page.
See the benchmarkBuilt on NVIDIA OpenShell; no NVIDIA endorsement implied.
FAQ
Two things. Datasets, evals, and RL environments for frontier model labs. Small models for enterprises. Both come from traces our hypervisor and W8 browser record.
Live traffic is captured at the model proxy. That record holds prompts, completions, and the tool calls the model saw. The hypervisor adds the processes, network calls, each allow or deny, and the file changes behind each call. W8 adds the whole page at every step. So the training set holds what the agent did and what the page held, next to what the model said.
A recording of an expert doing a real task inside our hypervisor and W8. It holds each action, each runtime event, and the page at each step. An agent trace is the same record made by an AI agent.
No. You keep your agent's logic and add no tracing SDK. Setup covers the image, the sandbox config, keys, and gateway routing.
No. Enterprise data and lab data sit under separate contracts. Any license to a third party needs explicit rights and a stated use.