Project summary

As AI agents start taking real action on our computers like using tools, running skills, and handing workflows, it is very easy to lose track of what they are doing. If an agent makes a mistake, users often have no easy way to stop it or undo the damage.
A los of AI safety ideas sounds good on paper, but they often fail when tested in real environments. My goal is to bridge that gap. I want to take theoretical concepts about agent identities, strict tool permissions, and action logs, and put the to the test inside a real desktop workspace environment.
I will study how these pieces actually work together when an agent tries to execute tools, where permissions fail, and how we can make agent changes completely reversible. By testing these ideas practically rather than just theoretically, I want to figure out what actually keeps humans safely in control, and share those real-world findings with the developer community.

What are this project's goals? How will you achieve them?

Goals:

1. To figure out why AI agent governance and safety rules often sounds good on paper, but break down when used in real software.

2. To test how agent identities, tool permissions, and undo mechanisms actually interact inside a working environment.

3. To discover what practical safety measures are truly needed to keep humans in control.

   
   How I will achieve them:

   1. I will analyze the gap between theoretical AI safety models and real word execution.

   2. I will use a desktop workspace environment as a testing ground to run hands on experiments with agent tools, skills, and permission limits.

   3. I will document where the safety barriers fail, see how permissions and undo features hold up under real tasks, and share what I learn so others can build safer systems.
      

How will you know it worked?

1. Clear findings showing the exact points where agent permissions, tool limits, and safety controls break down during real usage.

2. An AI agent history log that clearly showcases real AI agent behavior, tracking what steps they took, what tools they used, and how they handled permissions during task.

3. Practical insights on what makes agent changes reversible and how human can retain control.

How will this funding be used?

The funding will support the research process and cover basic living expanses so I can focus full time on this work:

1. Minimal living costs: covering basic living expenses to allow for dedicated, full time focus on the research without external distractions.

2. Access to AI models: paying for API usage and subscriptions for leading AI models needed to test agent behavior and capabilities.

3. Access to research: access to the best scientific papers, academic journals, and technical resources on AI safety and governance.

4. Access to tools and courses: access to the software tools, development environments, and related educational courses to support the investigation.
   

Who is on your team? What's your track record on similar projects?

I am an independent researcher and developer with deep, hands on experience building applications featuring AI governance and agentic workflows since the early days of modern AI.
I have a strong foundation in IT technologies, holding Master's degree with top honors for advanced research in the field.
I have been actively involved in building agentic tools and exploring how AI systems interact with software environments. This practical background gives me insight in to the real world challenges of agent safety, permissions, and tool execution that this research aims to solve.

What are the most likely causes and outcomes if this project fails?

In research, there is no true failure because testing the limits of safety is the goal.
The core question: the main thin being tested is whether current existing mechanisms that govern AI agents actually operate safely in practice or not.
The outcome: if current mechanisms break down or fail under real testing conditions, that is not a project failure, but it is a critical discovery. Finding out where and why current governance tools fail gives us the exact data we need to figure out what safety measures must be fixed or build next.

How much money have you raised in the last 12 months, and from where?

I have not tried to raise money previously. I am self funding up to this point, and I only recently realized that securing dedicated research funding would make it possible to perform this work professionally and full time. This is my first attempt to get external funding.