AI agents and automation, with a human in control

Who it's for

• Teams that want an agent to take over a repetitive task but need to approve what it does.
• Products that need a voice or chat assistant grounded in real facts, not a demo.
• Anyone who wants to test whether an agent can do the job before paying for a build.

What you get

• Agents that ask before they act. In Eight Keys, agents ask permission on a physical key, show their progress, and pause or stop when I press.
• Natural language to one concrete action. Talking Vacuum turns a sentence in Hebrew or English into one action on a robot's real floor plan.
• Assistants that stick to facts. The AI twin on this site talks, listens and gives voice-guided tours, and answers only from a facts file.
• A test before the build. Every tool starts with a cheap feasibility test on real inputs and on decoys where the right answer is nothing.

How it works

1. Name the task and the decision points where a person has to approve.
2. Test the premise cheaply on real inputs and on inputs where the answer should be nothing.
3. Build the agent and its controls together: approval, progress, pause and stop.
4. Hand over with the facts it is allowed to use written down.

Questions

How do you stop an agent from doing something it shouldn't?

By putting the decision points in front of a person. In Eight Keys every risky action waits for a physical key press, and any running agent can be paused or stopped.

Will the agent make things up?

It answers from a defined set of facts. Before trusting a tool I test it on inputs where I know the right answer is nothing, and I make it say when it is guessing.

Does it work in Hebrew?

Yes. Talking Vacuum and the AI twin on this site both work in Hebrew and English.

Do you share what you learn?

Yes, in long-form posts under Writing, including the experiments that failed, like the shared library four agents never used.

Want to talk about your team? A two-line email is enough: ilan@ilans.net.