The Agentic Web Manifesto

It started with frustration.

I was building with AI agents, and every new session felt like starting over. My agent had no memory. It didn't know my stack, my patterns, or the decisions we'd made yesterday.

The context was there, somewhere. Trapped in Cursor's cloud. Scattered across chat histories. Locked in proprietary databases I couldn't access.

I kept asking: Why doesn't my agent remember?

The Real Problem

Every AI coding assistant today operates in a silo. Cursor doesn't talk to Copilot. Claude doesn't know what you told Windsurf. Each tool builds its own walled garden of "context" that never leaves their servers.

Meanwhile, your codebase— the one thing all these tools share— has no memory at all.

We're building the internet of agents without a protocol layer. No shared memory. No standard format. No way for agents to pass knowledge to each other.

What If Context Lived in Your Repo?

I started experimenting. What if agent memory wasn't stored in the cloud, but right next to the code? A folder. Some markdown files. A simple structure that any agent could read and write to.

That experiment became UCP (Unified Context Protocol)— a file-based standard for persistent agent memory.

But here's the thing: UCP is just my answer. There are thousands of developers out there with their own systems. Markdown collections. Prompt libraries. Custom workflows. All solving the same problem in different ways.

Solving the "Knowledge Gap"

There's another problem: The Training Cutoff.

Scenario: Supabase releases a new Auth API today. Your AI Model (trained months ago) doesn't know it exists. It hallucinates old code, and you spend hours debugging.

CorePack AI acts as a Knowledge Patch. When you install @supabase/auth-v2, you are essentially "firmware updating" your agent. It reads the new rules from disk and instantly becomes an expert in a library that didn't exist when it was trained.

Installable Intelligence

Here's the vision: You want to add Stripe integration to your project. Instead of reading docs and copy-pasting code, you run one command. Your agent instantly knows the best practices, security patterns, and common pitfalls.

What Can You Build?

We're calling these "Context Packs." Every industry, every domain, every specialized workflow— there's an opportunity to package that knowledge.

The "Senior Engineer" on Disk

In a team, a Junior Dev might say "Make a button" and get raw HTML. A Senior Dev says "Make a button using our design tokens" and gets a robust component.

With CorePack AI, the Senior Dev installs @company/design-system once. Now, when the Junior Dev asks for a button, the agent checks the rules and automatically generates senior-level code. The prompt drift creates consistency.

Beyond Code: Seeds of Innovation

The protocol is text-based, which means it works for any knowledge work. We envision a future where every profession has its own "Senior Expert on Disk."

Our First Pack: UCP

To kickstart this ecosystem, we're releasing UCP as the first official Context Pack. It's opinionated. It's strict. It handles the hard stuff— state persistence, multi-agent coordination, session handoffs.

But UCP doesn't have to be your answer. CorePack AI is protocol-agnostic. Bring your own system, share it with others, or use ours.

About the Author

Nikhil Rao

Building tools that bridge human intent and autonomous execution.

X (Twitter)LinkedInLinktree