Windows-first · for developers, operators and technical teams
Governed execution infrastructure for AI agents on real computers.
PBT-G is a controlled execution layer for AI assistants that need to work across browser, terminal, file and local workflows — with explicit boundaries, supervised operation and reviewable activity when AI moves from recommendations to actions.
Product introduction
See the product story in under two minutes.
A concise introduction to the problem PBT-G addresses, the operator-control model and the kinds of supervised browser and computer workflows the product is designed to support.
How it works — conceptually
Every action passes a boundary the operator controls.
Pick a surface and run the simulation. The request reaches the boundary, waits for an operator decision, and leaves a record either way. This is an illustration with fictional data, not a product screenshot.
Supervised browser tasks — navigating, reading and working in web pages — within the sites and actions an operator has allowed.
- AI requestAn assistant asks to act on the computer.
- Governed boundaryDemo scope: approved project sites, read and summarize
- Authorized actionOnly an authorized action proceeds.
- Reviewable recordThe outcome stays reviewable afterwards.
Simulation ready.
Run the simulation to see a request move through the boundary.
Why it matters
Agents are learning to act. Real computers need a boundary.
AI assistants increasingly do more than answer — they click, type, run commands and move files. On a real computer, those actions touch real work and real accounts.
PBT-G focuses on the space between agent autonomy and operator control: making agents useful on the machine without making their access unlimited.
Agent autonomy
- Initiative and speed on routine work
- Multi-step tasks across tools
- Fewer manual hand-offs
Operator control
- Clear limits on what can be touched
- A human decision where it matters
- Activity that can be reviewed later
PBT-G is designed so teams do not have to choose one over the other.
Capabilities
One governed path for the places agents actually work.
- Browser
Supervised browser tasks
Designed for agents that need to navigate, read and work in web pages while an operator stays in control of where they go and what they do.
- Terminal
Controlled command execution
Built to support terminal work inside explicit limits, so useful automation does not require handing over an unrestricted shell.
- Files
Scoped file work
Focuses on letting agents read and organize files only within the locations an operator has authorized.
- Local workflows
Multi-step local work
Built to support tasks that span browser, terminal and files on a real Windows computer — not just a chat window.
- Governance
Operator-set boundaries
Operators decide what an agent may do. Work outside that scope is designed to wait for a human decision.
- Reviewability
Reviewable activity records
Agent activity is designed to leave a clear record that people can look back on when they need to understand what happened.
Use cases
Built for supervised, reviewable agent work.
Supervised browser tasks
Research, form work and routine web tasks carried out by an agent, with the operator governing the boundary rather than watching every click.
Controlled local workflows
Repeatable desktop and developer chores — preparing files, running checks, assembling reports — kept inside an authorized scope.
Auditable agent operations
For technical teams that need to answer “what did the agent do?” with a record instead of a guess.
Human-governed automation
Automation that moves quickly through routine steps and pauses for a person on the steps that matter.
Trust principles
Practical governance, stated plainly.
- Least privilege
- Agents get the access a task needs — not the access a machine happens to allow.
- Explicit boundaries
- What an agent may touch is stated up front, not inferred after the fact.
- Operator control
- A person can approve or decline agent actions and stays accountable for the outcome.
- Reviewable activity
- Activity is designed to be easy to inspect after it happens.
Complements what you already use
PBT-G is designed to work alongside existing AI assistants and to complement existing endpoint security and isolation practices. It does not replace OS-level sandboxing, endpoint security or organizational security controls.
Status
Early-stage and in active development.
PBT-G was founded in 2026. It is not yet generally available, and the descriptions on this page reflect the product’s design goals and direction.
- Now
Windows-first controlled execution for browser, terminal, file and local workflows.
- Next
A clearer operator experience for approving actions and reviewing activity.
- Later
Governance that works for technical teams, not only individual operators.
Direction, not commitments. Priorities may change as the product develops.
Early access
Evaluate PBT-G with your workflow.
PBT-G is not yet generally available. We are selectively onboarding developers, technical teams and early design customers who want to evaluate governed AI execution on Windows.
Start with a scoped conversation about your workflow, the surfaces an agent needs to use and the boundaries your operators need to keep. Access and commercial terms are discussed case by case while the product remains in active development.
Company
PBT-G
Open to conversations with technical teams, partners, startup programs and investors interested in governed execution for AI agents.
- Founder & Product Lead
- Senih Bayankulu
- Location
- Izmir, Turkey
- Founded
- 2026
- Focus
- AI · Cyber Security · Software Development · B2B / Enterprise