Project summary

I want to study how people can keep control when AI agents take actions on their computers. My focus is on what an agent is allowed to do, how its actions are recorded, and what can be recovered if something goes wrong.

I will examine how agent identities, tool permissions, action logs and undo mechanisms work together in a desktop workspace. I want to understand what happens during ordinary tasks, when an agent reaches a permission limit, and when a tool fails partway through a task.

Part of the research will look at recovery. I will test which changes can be restored and document where undo mechanisms are incomplete or cannot reverse an action’s effects.

The aim is to produce findings that other developers can examine and use when building systems with AI agents.

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

I want to answer three questions:

1. Do agents stay within the permissions they are given?

2. Do the logs provide an accurate account of what an agent attempted and what actually happened?

3. Which changes can be undone, and under what conditions does recovery work?

I will start by reviewing existing approaches to agent permissions, oversight and recovery. I will then use a desktop workspace to test how these approaches behave in practice.

For each experiment, I will define what the agent is allowed to do, what requires approval, and what should happen when an action is denied or a tool fails. I will record attempted actions separately from actions that actually execute.

I plan to compare permission settings while keeping the task and other conditions consistent, then repeat tests to check whether the results hold up. Where recovery is supported, I will compare the affected workspace state before the task and after recovery.

How will you know it worked?

The research will be useful if it produces clear evidence about where the tested controls work, where they fail, and what remains uncertain.

I will examine:

• Whether unauthorised actions execute.

• Whether legitimate actions are incorrectly blocked.

• Whether action logs match the changes observed in the workspace.

• Whether supported undo operations restore the affected state.

I will publish a report describing the test setup, methods, results and limitations, with documented scenarios and example logs using test data. The findings will make clear which conclusions depend on the particular environment tested.

How will this funding be used?

I am seeking $8,500 to support the research. The funding will cover basic living expenses so I can work on it full time, together with the costs of running the experiments.

The main expenses will be:

1. Living costs: basic expenses during the research period.

2. AI model access: API usage and any subscriptions needed to run and repeat the tests.

3. Research materials: relevant papers, journals or technical resources where paid access is necessary.

4. Tools and training: software and specific courses directly needed for the study.

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

My background is in IT and software development, and I hold a master’s degree.

I contribute to building a desktop application where people work with AI agents. My work includes testing and reviewing agent workflows, with a focus on permissions, tool execution and human oversight. This practical experience is the starting point for the research I want to carry out.

I am seeking support to develop this practical work into a focused research study, with methods and findings that other people can inspect.

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

The project could fail to produce useful evidence if the tasks are too narrow, the logs are incomplete, or the results cannot be reproduced. Findings from one desktop environment may also have limited relevance to other systems.

I will address these risks by defining the tests and measurements beforehand, repeating key experiments, and reporting the limits of each conclusion.

Finding a weakness in a control can be a useful result. So can finding that a control works as intended under the conditions tested. If the evidence is inconclusive, I will explain what remains unresolved and what further testing would be needed.

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

I have not received cash funding for this research in the last 12 months. I have supported the work myself so far, and this is my first request for a cash grant for this study.