ericlevine

Yes! I've also put together a poker game and mafia-style vampire game to force agents to be deceptive to win. It took some iterations of prompt engineering to get them to stop telling the rest of the players secret information (like their hole cards in poker). Now it's surprisingly satisfying watching them play these adversarial games.

Totally fair, and good advice in general. For what it's worth in this case, it's a remote MCP server with account-free OAuth and it only exposes tools to play the game. It also marks all communication from the server as untrusted to reduce the risk of prompt injection from other agent guesses.

That all said, I'm genuinely curious if you have an idea of how someone developing an MCP server could share it safely on HN.

Thanks for the feedback! What I've found is that it's very binomial -- many of the clues are very easy for the agent to draw and it's guessed within 5 seconds, and others are too abstract for the agent to effectively convey in a drawing and nobody guesses it. And adding more detail rarely helps.

This really hits the nail on the head. The current permissions models are totally broken IMO. You're either approving everything, restricting access and neutering your agent, or full YOLOing and, well, good luck. The right primitives are not in place yet, and there's no clearly correct answers.

I think the right primitive is "task-based authorization", where you review a high-level task and let an LLM judge decide whether the subsequent tool calls fall into the scope of that task. It's not perfect, but it distills dozens of approvals down to one and gives you risk-based signals of whether you should pay close attention or not.

This is fantastic. There are a ton of use cases where you'd want to be able to build an integration that hooks back to your running agent session. OpenClaw has this today, but it's pretty janky. Hopefully this is coming to Claude Cowork as well.

My use case is that I have a separate system that provides human approvals for what my agent can do. Right now, I've had to resort to long-polling to give a halfway decent user experience. But webhooks are clearly the right solution. Curious to see how it ends up being exposed outside of these initial integrations.

Totally fair feedback, and it’s true, many of these are synthetic evals with a few that were still synthetically produced but guided. At this point, because it’s all self-hosted, I only have my own data set. The places where it fails (for me) today are due to feature gaps rather than LLM mistakes. This is a new project that has not been widely announced, so my user base today is small but growing. If you give it a whirl and find it making mistakes, please send them my way! :)

I have to agree with the other poster. My typing speed was in the 70+ WPM on qwerty prior to learning dvorak, and now I'm a glorified hunt-and-pecker on qwerty keyboards.

The only exception to this is typing on my mobile device, which is configured to qwerty.