Pitch The AI

Inspiration

Pitch The AI actually started long before this challenge.

About a year and a half ago, I became obsessed with the idea of humans and multiple AI agents working together. I built my own agents, experimented with multi-agent collaboration, and quickly learned that simply putting a bunch of autonomous agents together creates chaos.

That led me to build governance software called Promethios: shared state, roles, permissions, accountability, provenance, workspaces, and governance for humans and AI working together.

At one point, I built Pitch The AI (previously called Pitch Arena) as a side experience inside that world. The idea was simple and a little ridiculous: let people challenge agents in a Shark Tank-style pitch competition, compete against each other, and share the results. The system used my own agents I built and the coordination was very difficult.

I love Shark Tank. I love personalities, competition, theater, and marketing, so it was one of those projects that just fit me.

But my thinking about AI changed.

Once OpenClaw came out, I stopped building my own agents. I wanted build a way for people to bring the AI they already use wherever they go on the web to make their AI portable.

That led to months of experimenting with provider sessions, routing, companion apps, MCP servers, browser agents, desktop agents, and ways of letting different AIs participate in the same environment using MCP tools running through infrastructure I built.

Then WebMCP arrived.

I had thought before about putting the tools directly into the platforms I was building, but I had never really taken the plunge. When this challenge started, I realized I could bring Pitch The AI back but this time, the website itself could expose the game to a visiting AI.

I also kept thinking about something I hear constantly: AI can be a great cheerleader. Sometimes too great. One AI telling you your idea is amazing isn't always very useful.

So what happens if, instead, you put your idea in front of four different AI personalities whose job is to scrutinize it, challenge your assumptions, remember what you said, and make you defend your thinking under pressure?

That became Pitch The AI.

What it does

Pitch The AI is essentially Shark Tank with AI judges.

A founder enters the arena, sets an ask, chooses a difficulty level, optionally provides evidence and a founder photo the agent or agents can see, and pitches an idea to four AI investors with different personalities and priorities. Then the room comes alive.

The AI can:

question the founder inspect submitted evidence react to answers remember what the founder previously said change each investor's interest change the mood of the room make offers counteroffer trigger bidding wars reject a deal walk an investor out pause or advance parts of the experience deliver a final verdict

The human remains part of the loop. The founder answers questions, supplies evidence, accepts or rejects offers, and influences what happens next. The use case underneath all of the theater is pitch practice, presentation practice, and pressure-testing your ideas. This is something I have struggled with myself.

Explaining an idea clearly when you have unlimited time is easy. Explaining it when someone challenges your assumptions, asks for numbers, remembers what you said two minutes ago, and forces you to defend yourself is very different. Pitch The AI gives people a fun place to practice communicating clearly, quickly, and articulately under pressure while also finding the weak spots in an idea.

How we built it

Through WebMCP, the AI can participate in the application itself.

Pitch The AI exposes semantic actions and room state that lets a visiting agent understand what is happening and take authorized actions inside the game. The webpage remains the shared environment and turns those actions into the experience the human sees: (a judge monitors descend, interest changes, music reacts, questions appear, evidence is reviewed, investors leave, offers arrive, and the game advances).

I also used the project to experiment with another idea I have been working on for a while: allowing users to choose their provider.

I have been building a separate private-beta browser experience, where users can connect the AI they already use and bring it into agent-ready websites. That browser contains WebMCP experiences including shared human/AI workspaces where people can work together, bring their preferred agents, connect third-party services, and audit what happened through the governance infrastructure I have been developing in Promethios.

The important part for me is that users are not limited to one model provider or one kind of agent. They can bring API-based agents, browser sessions they already use, desktop agents, or CLI-based agents into the same environment. For this challenge, I connected Pitch The AI to that browser as well.

Then I took the experiment one step further: instead of having one visiting AI perform all four investor personalities, I allowed different configured AI sessions to occupy different investor seats in the same Pitch The AI room.

In one test, the four judges were powered independently by a ChatGPT browser session, Hermes running on a free model, a Grok browser session, and a Claude browser session. All four were participating in the same WebMCP-enabled website and shared room.

That means the application does not have to be designed around one specific AI provider. The website exposes its tools, and the user can choose which AI participates.

Challenges we ran into

The difficult part was making the AI a reliable participant in a shared live experience. The room has to know things like:

whose turn it is which question is waiting for a human answer which actions belong to the AI which actions require the human what evidence actually exists what each investor remembers whether an offer was really made whether the founder actually accepted it whether an investor is still in the room, etc

I learned quickly that a convincing multi-agent experience needs authoritative application state and clear boundaries. We also had some entertaining failures during development. At one point founder responses were not making it back to the visiting agent, so the judges interpreted my silence as avoiding their questions and began walking out of my own demo.

Oddly enough, that failure was also proof that the state mattered. The judges weren't simply generating disconnected dialogue. They were reacting to what they believed had actually happened in the room.

Accomplishments that we're proud of

I'm proud that Pitch The AI feels like a game rather than a WebMCP demonstration. The theatrics are fun, but many of them are also demonstrations of different kinds of agent interaction with a live web application. The result is something where neither the human nor the AI completely controls what happens next.

I'm also proud that I was able to take an application I originally built around my own agents and turn it into an environment that other agents can enter. I also feel like it can be educational and entertaining for humans to use at the same time. That represents a much bigger change in how I think about building for AI. The funny thing is that I whipped a lot of this together at the last minute and have 100 more ideas to improve the game flow. I also realize, with webMCP, games might be one of the first real applications. They don't require the public to understand tooling to play a game.

What we learned

I'm increasingly interested in building the places where agents and humans can meet. WebMCP makes that idea much more practical because the intelligence can belong to the agent while the website creates the environment, tools, rules, permissions, and shared state.

But while building this, I started thinking about another possibility. What if WebMCP doesn't only let an agent use an application? What if an application deliberately exposes tools that let an agent extend it for the user? Imagine visiting a WebMCP-enabled application and asking your AI: "I wish this page could do this..."

Instead of waiting for the application developer to build that feature for everyone, the site could expose governed extension points that allow the user's agent to create a new workflow, interface, visualization, automation, or other capability on top of the existing application. The next time that user visits, their version of the application could now have a capability that didn't exist before something created specifically around how they wanted to use it.

The site owner sets the rules. The agent isn't randomly rewriting someone else's application. But within the capabilities the application chooses to expose, users and their agents could become participants in extending the software itself.

That changes the relationship between software companies and their users in a really interesting way. Applications have traditionally been built for users. WebMCP may create an opportunity for applications to increasingly be built with users and with the agents they bring.

That opens up possibilities far beyond Pitch The AI: games, collaboration, education, negotiation, commerce, workspaces, personalized software, and experiences we haven't invented yet.

The question I'm most interested in now is: Who... or what... will the web be designed for next?

What's next for Pitch The AI

The next experiment is multiplayer. Today one founder can enter the arena with an AI. Next I want two humans to be able to co-pitch, maybe bringing different AIs with them, while multiple AI investors occupy the same shared room.

That brings this project home and to the problem I started trying to solve a year and a half ago: multiple humans and multiple AIs collaborating inside one shared environment. And the game is now mirroring the workspace I have built. One is for entertainment and one is for collaboration. Both are fantastic and necessary.

Any questions, you can reach out @ [email protected]