Instant agents, portable skills, and why architecture matters

2026-05-24

by Uri Walevski

When we launched prompt2bot, we wanted to make building AI agents feel less like software engineering and more like writing. But the bigger shift is in how we think about capabilities. We don't think you should have to build a new bot every time you want to try a new task.

That is why we built Talk to Skill. You point at any skill from Tank, npm, or GitHub, and we instantly spin up a personal AI agent equipped with that exact capability, ready to chat.

The Problem with Heavy Infrastructure

In most agent setups, standing up a new bot is a heavy operation. You have to configure databases, set up servers, handle OAuth flows, build a UI, and wire up messaging channels. If you want to try three different skills, you end up building three different software systems.

The alternative is the monolithic agent, where you load every tool and guide into one massive prompt. But as we've written before, that just leads to tool confusion and prompt bloat. Your agent spends more time trying to navigate its own system instructions than actually helping you.

To make skills portable, the infrastructure has to be cheap. If spinning up a bot costs five dollars a month in fixed server hosting, you can't have a hundred ephemeral specialists. You have to consolidate.

Why prompt2bot is fast and cheap

Our architecture is built on a different set of assumptions. Every piece of the stack is designed to be stateless and ephemeral.

First, the agent run loop lives entirely on serverless edge functions. There is no persistent Node.js process waiting for messages. The code only executes when a message arrives and shuts down the millisecond the response is delivered.

Second, our database and state are serverless. We use highly reactive, serverless state and history storage. We don't pay for idle database servers, so you don't have to either.

Third, our execution environments are provisioned on demand. When an agent needs to execute code or manage virtual machines, we spin up an ephemeral sandbox or a lightweight virtual machine. That environment lives for the duration of the task and gets destroyed automatically after one hour of inactivity.

Because the cost of an agent collapses to the exact compute used during active runs, we don't have to charge you to keep bots warm. We can let you spin up a brand new, private specialist in one click, for free.

How to use portable skills

This architecture makes skills feel like software packages. You can browse tankpkg.dev or GitHub, find a capability you want, and launch it instantly.

For example, pointing to tank:@prompt2bot/therapist creates a private AI therapist on WhatsApp. You can even add it to group chats to act as a couples counselor. The therapist skill itself was built entirely by our Builder AI using another skill, the Tank Skill Creator.

Our default templates work the exact same way. If you want a coding assistant or a personal assistant, you can launch them directly via p2b-coder and p2b-personal-assistant.

If you need something highly specific, you can just ask the Builder AI to write a new skill for you. You describe the API, the database, or the tasks you want, and the builder handles the coding, testing, and publishing to Tank in the background.

Monetizing your skills

Once you have a skill that works, you can share it. And if you are logged in when you generate a skill link, we automatically append your affiliate reference.

If anyone clicks your link, spins up an agent, and eventually decides to upgrade to a paid tier to unlock more VM time or agent runs, you earn a 20% recurring monthly commission. You don't have to set up Stripe, manage subscriptions, or support servers. You write the skill, we handle the infrastructure, and you get paid for the value you share.

The pieces are all live on the Talk to Skill page. Try launching a template, ask the builder to write a custom skill, or grab your referral link and start sharing.

← All posts