This is not a hypothetical. This is what happened when I gave Perplexity Computer a real-world task, my login keys, and a few minutes of spoken notes.

For the past two years, we have learned to talk to AI. We ask it questions, and it gives us answers. Perplexity, once known for its smarter web search, has evolved. It's no longer just a consultant; it's becoming an employee. It can not only offer advice but also do the work.

That era is now ending.

This week, I ran an experiment that convinced me we have crossed a significant threshold. I wanted to create a new microsite for my AI Toolbox — formerly a poorly formatted page on my Beehiiv website. Instead of hiring a developer or spending a weekend wrestling with code myself, I gave the task to Perplexity Computer.

I did not type a detailed prompt. I spoke for a few minutes into Wisprflow, a voice-to-text tool, outlining my vision for the site. I also had the legacy page with all the tools I liked and was evaluating. Then, I gave the transcribed notes to Perplexity Computer and asked it not only to build the site but also to make the core architectural decision: should this be a custom application or a WordPress site?

Then, I did something that felt both futuristic and reckless: I gave it the keys. It hurt, but I broke out the credit card and signed up for a month of Perplexity Pro for $200. I granted Perplexity Computer access to my GitHub and Vercel accounts and let it run.

What happened next was nothing short of impressive.

This AI Lesson will walk you through that exact workflow. We are moving beyond hypothetical tutorials to show you what happens when you delegate a real, non-trivial project to a persistent, autonomous AI agent.

Skill Level: Advanced — This workflow involves granting AI access to production services such as GitHub and Vercel. Proceed with extreme caution and at your own risk.

AI LESSON

What happened when I gave an AI agent a real-world task, my login keys, and a few minutes of spoken notes.

This is not a theoretical exercise. This is a real project. The goal was to build a microsite for the AI Toolbox, a curated list of AI tools I recommend. The process demonstrates a new paradigm for work: human-led, AI-executed. Here’s the end result.

Instead of writing a detailed project brief, I used a voice-to-text app, Wisprflow, to capture my thoughts. I spoke for about two pages worth of notes, describing the site's purpose, the kind of content it would host, and the general feel I was going for. The app transcribed my rambling into a structured text document. This became the entire project brief.

This initial step is crucial. The quality of the AI output is directly tied to the quality of your input. Using a tool to structure your thoughts — whether it is voice-to-text, a mind map, or a simple bulleted list — is the foundation of successful delegation.

My prompt to Perplexity Computer was fundamentally different from a standard AI query. I did not tell it what to build. I asked it to decide.

My prompt looked something like this:

This is the key shift in how to work with a tool like Perplexity Computer. I was asking the AI to act as a CTO, not just a coder. Perplexity Computer analyzed the request and came back with a nuanced recommendation: build the site using Next.js (a JavaScript library) and host it on Vercel (a low-cost, high-quality application hosting provider) for performance and scalability. Crucially, it also suggested that a WordPress backend could be integrated later for easier content administration if needed or the AI turned out to be a less than capable webmaster. This was not a simple either/or answer; it was a strategic, phased approach that balanced immediate needs with future flexibility.

Perplexity Computer requires a Pro subscription (currently available to Perplexity Max subscribers; enterprise availability is coming soon), which runs $200 per month at the enterprise tier, or $20/month for the standard Pro plan. I signed up for a month to run this experiment — and yes, it stung a little. But the question I was asking myself was not whether $200 was expensive. It was whether $200 was cheaper than a developer's hourly rate for the same work. It was not even close.

Once I approved its plan, Perplexity Computer asked for access to my GitHub and Vercel accounts. This is the moment that gives IT departments nightmares, and it should not be taken lightly. I used OAuth to connect the services — which is more secure than sharing passwords — but it still grants the AI significant permissions to create repositories, write code, and deploy to production.

Perplexity Computer then:

A note on trust and access: Best practice is to create dedicated service accounts or use platforms that support granular, token-based permissions for this kind of automation. Before granting any AI agent access to production services, understand exactly what permissions you are granting and whether they can be scoped to the minimum required for the task. **This is not a toy. You are granting access to your production systems. Proceed with extreme caution.

Every Day Updates with Manus

I am not sure I will use Perplexity long term, but I wanted to make the website easier for me and my admin, Christine, to update. Also, my Perplexity results weren’t perfect. It seemed to get stuck making updates. So I created a project in Manus, my go-to desktop agent, and added custom instructions and a style guide. Now, whenever I need to make a tweak, I don’t code; I just describe. For example, on the original website, I felt like there was too much vertical space. I asked it to shrink the vertical space of the hero image, and it did it. I wanted to generate a sitemap.xml for search, I asked, and I received.

The entire process — from voice note to live website — was handled autonomously. Perplexity Computer made the right architectural choice, wrote clean and functional code, and managed the full deployment pipeline without any intervention from me. But it did get stuck while trying to upload the site logo. I ended up using Manus to fix that.

It was not just a tech demo. It was real work, completed by an AI agent that had access to the right tools, a clear goal, and the authority to make decisions. The result was a professional, functional microsite, delivered without me writing a single line of code or configuring a single server. The AI was able to take my unstructured notes and turn them into a complete, working application. This is a powerful new capability that will change how we build software.

As impressive as this is, it’s important to be realistic about the current limitations. These AI agents are powerful, but they are not magic. Here’s what this workflow still struggles with:

This workflow is a template for a new kind of professional leverage. You no longer need to be a developer to build custom software. You need to be a good project manager who can clearly articulate a goal and then delegate it to a capable digital worker.

The real shift here is not just about building websites. Think about what this capability means at scale. For example, you could feed an AI agent the transcript of a client discovery call, along with your internal notes on the project, and ask it to generate a formatted, professional-looking Statement of Work (SOW). The agent could extract the key deliverables, timelines, and pricing information and structure it into a customer-ready document, saving you hours of administrative work. You can describe the presentation you need, and it will generate a complete slide deck with custom graphics and on-brand design. You can describe a data problem and have it write the code to solve it. The pattern is the same: structured input, delegated execution, human review.

Here is how to apply this pattern to your own work:

We are at the very beginning of this shift. The tools will improve, the safety guardrails will mature, and the scope of what we can delegate will expand. But the opportunity is already here, and it is already impressive. The question is no longer whether AI can do real work. The question is what real work you are willing to give it.

I appreciate your support.