Skip to main content
W8 Intelligence

The right AI
is built in-house.

Ground truth from your own work. Our team embedded in yours.

The hypervisor and W8 record the work your people and agents do today. Our engineers sit with yours. Together we turn checked runs into evals, environments, and a model that runs inside your walls.

Talk to usHow we work
01Compliance

Compliance and security are how the intelligence becomes yours.

Agents act on live systems. So a review asks three things: where the rules are enforced, where the keys live, and what record survives. Answer them below the harness, and the record that passes the review becomes the training set for a model you own.

The challenge

Agents act. Their guardrails and their records live in the wrong place.

  1. 01

    The risk moved from answers to actions.

    An agent writes code, calls APIs, and changes live systems. A review has to account for what ran, where it connected, and what it changed.

  2. 02

    Guardrails written in the prompt.

    The next prompt can overwrite a rule in the prompt. One injection or one bad tool is enough to walk past it.

  3. 03

    Credentials inside the agent.

    Keys in the agent's environment travel with everything the agent reads. A leak needs no exploit, only a request.

  4. 04

    A record the agent writes about itself.

    The harness logs what its vendor chose to log. A model proxy keeps what the model said. Nobody keeps what ran, what was refused, and what the page held.

  5. 05

    Training on trust.

    Runs enter training with no record a reviewer can open. The model inherits each gap.

  6. What a review asks for

    Isolation. Rules enforced where the agent cannot reach them. Keys the agent never holds. A runtime record of what ran and what was refused.

The offer

One record, from control to model.

  1. 01Control

    Each action passes rules enforced outside the agent, below the harness.

  2. 02Record

    Each run keeps its evidence, including what was refused.

  3. 03Model

    Only approved runs with evidence enter training. The weights come from a record your reviewers can open. They are yours.

The chain runs through one record, kept below the harness. A model proxy keeps what the model said, and that is where it stops. We are the only ones who turn that record into a model you own.

02How we work

Embedded with your team, from the first eval to the model in production.

Our engineers join at the first eval and stay until the model ships. Each engagement goes live and gets measured there.

  1. 01

    Discover

    Pick one workflow with your team. Define the first eval and the outcome it scores.

    WootzApp team
    Your team

    Names the workflow and the experts who define a correct run.

    WootzApp

    Writes the first eval with those experts.

  2. 02

    Capture

    Your harness runs inside the hypervisor and W8. Agent runs and expert work become ground truth.

    HypervisorW8 browser
    Your team

    Runs the workflow as it does today.

    WootzApp

    Installs the hypervisor under your harness and W8 behind your browser agents.

  3. 03

    Train

    Checked runs become evals, environments, and a small model trained on your work.

    WootzApp team
    Your team

    Approves each run that enters training.

    WootzApp

    Builds them and runs the training.

  4. 04

    Run

    The model serves inside your walls, with frontier fallback at your gateway. New checked runs keep improving it.

    HypervisorW8 browser
    Your team

    Approves each model that ships and owns the gateway.

    WootzApp

    Hands over the runbooks and stays for the retraining rounds.

03Ownership

You own the intelligence.

Your edge is the work your people do today. A model trained on that work belongs to you.

Your data

Recorded for your model alone. Enterprise data and lab data sit under separate contracts.

Your model

The weights, evals, and environments built from your runs are yours to run and improve.

Your perimeter

The model serves inside your walls, with frontier fallback through your gateway.

04Workflows

Workflows where the record exists today.

Illustrative

Any task your agents or experts run inside a sandbox or a browser is a candidate.

  • Back-office operations

    W8 browser
    Example work
    Claims, invoices, and account matching in browser tools.
    The record
    The exact value on each screen, even content off screen.

  • IT operations

    Hypervisor
    Example work
    Tickets, scripts, and config changes.
    The record
    Each process, file change, and network call.

  • Software delivery

    Hypervisor
    Example work
    Coding agents in your repos.
    The record
    Compiler and test results as the grader. See the preprint →

  • Research and compliance desks

    W8 browser
    Example work
    Experts gathering evidence across internal and public sites.
    The record
    The pages visited and the values read, as the eval and the training set.

These are examples. Each engagement starts with the workflow your team picks.

05Security

Answers for the security review.

The hypervisor is open source. The policy, the sandbox boundary, and the record it keeps are there to read before the first call.

For the review04 answers
01Where does the model run?
In your VPC, on your premises, or on servers we run. In each case the model serves inside your perimeter.
02Who can reach the sandbox?
Sandboxes expose no SSH port on the network. Sessions enter through the gateway with tokens that expire. Gateway traffic uses mutual TLS.
03Who approves what ships?
Your team approves which runs enter training and which model ships.
04What does the record hold?
Each run keeps its processes, file changes, network calls, allow and deny calls, and page states. Each reward points to its evidence.
Read the runtime architecture

Built on NVIDIA OpenShell; no NVIDIA endorsement implied.

06Questions

Questions.

Do we need to change our agent?

No. You keep your agent's logic and add no tracing SDK. Setup covers the image, the sandbox config, keys, and gateway routing.

What enters training?

Only checked runs. A run enters training when its evals pass and your team approves it.

What happens when the small model falls short on a request?

The gateway routes that request to a frontier model. The run is recorded all the same and joins the pool for the next training round.

How is this different from training on my live traffic?

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.

Do you resell enterprise data?

No. Enterprise data and lab data sit under separate contracts. Any license to a third party needs explicit rights and a stated use.

07Research

Built on research we publish.

Our benchmark, dataset, runtime, and training study are public. Each cell shows two figures from the entry and links to the write-up, data, and code. All work →

  1. 01Benchmark· Oct 2026
    Paired browser tasks
    100
    Audited rubric criteria
    642

    W8 DOM-diff Bench

    23.52% → 0.78% missing evidence

    Read the work
  2. 02Dataset· Oct 2026
    Tasks
    100
    Distinct sites
    90

    Browser Agent Tasks

    930 recorded browser states

    Read the work
  3. 03Preprint· Sep 2026
    C++ tasks in Fixed26
    26
    Task-trial runs per setup
    104

    Let Execution Do the Teaching

    45.2% first-round, 61.5% within two rounds

    Read the work
  4. 04Open-source runtime
    Project filesystem at /sandbox/work
    FUSE
    Sparse file chunks in PostgreSQL
    256 KiB

    Openrind Shell

    All /sandbox/work I/O through one daemon

    Read the work
The runtime

One boundary, decided outside the agent.

Your perimeterRecording, training, and serving
GatewayControl plane
  • Authenticates users
  • Issues per-sandbox keys
  • Delivers policy
  • Policy prover
SupervisorTrusted side of the boundary
  • Default-deny egress
  • Binary identity, SHA-256
  • Credential substitution
  • DNS and private-IP guard
  • Allow and deny log
SandboxThe agent runs here
  • Non-root process
  • Landlock file rules
  • seccomp syscall filters
  • Own network namespace
  • No credentials
  • One egress path

Your harness, unchanged. Each connection leaves through the supervisor.

01
Execution
Enforced in the kernel. Landlock limits files. seccomp limits syscalls. The agent runs as a non-root process in its own network namespace, with one way out.
02
Policy
Lives in the supervisor and the gateway. The agent cannot read or change it. Proposed rules pass the policy prover, and flagged changes wait for human review.
03
Credentials
Stay outside the sandbox. The agent sees a placeholder that the supervisor resolves only for approved hosts, ports, and paths.
04
The record
Each connection is checked by host, port, and calling binary. Each allow and deny is logged. The hypervisor keeps processes, file changes, network calls (denied ones too), and each allow or deny. W8 keeps the page at each step.
05
Training
Only runs your team approves enter training. Each reward points to its evidence in the record.
By extension, the hypervisor

WootzApp's hypervisor is OpenShell plus a project filesystem on a supervised FUSE mount. File I/O passes through one daemon under the same policy as the network. The record of a run holds both. Openrind Shell →

Built on NVIDIA OpenShell; no NVIDIA endorsement implied. The layers and controls follow the OpenShell architecture and security docs.

Start with one workflow.
Own the model that comes out of it.

Talk to us

W8 Intelligence. Built with your team, run inside your walls.