AgentGate

Inspiration

We set out to fix two fundamental gaps in today’s agentic systems.

First, most AI agents are locked into narrow use cases. The demos look impressive, but they don’t generalize well to real-world tasks. And realistically, navigating a browser is not an LLM’s preferred way to order an Uber. Instead of forcing a single general-purpose agent to do everything, we envision a network of highly specialized agents — each fine-tuned for a specific task.

Imagine asking your personal AI assistant to “book my favorite kind of night out.” Through Telegram, it understands your preferences, your occupation, your past behavior, and your local context. Instead of scraping random websites, it delegates to specialized travel, booking, and recommendation agents that return structured, high-quality results. Your assistant becomes an intelligent orchestrator — not a clumsy browser bot.

Second, there is no shared registry or standard interface for agents to discover and communicate with each other.

If you’re booking transportation across Europe, Lufthansa, SNCF, and Rome2Rio might each eventually have their own agent. But without a discovery layer and protocol, your personal assistant has no seamless way to interface with them. We’re building the infrastructure for an “internet of agents” — where companies can register their agents in a shared registry, making them accessible to everyone’s personal assistant.

What if the internet had “websites for agents” — and your local assistant could discover and talk to them safely?

That’s the future we’re exploring with AgentGate.

What it does

AgentGate introduces a modular agent ecosystem:

Local Assistant → Holds user trust and private context
AgentHub → Discovery layer for specialized agents
NetAgents → Task-specific agents (travel, legal, healthcare, etc.)
Visual Console → Transparency into delegation and execution

Instead of trying to do everything itself, your personal AI assistant:

Understands user intent
Discovers relevant NetAgents via AgentHub
Delegates tasks securely
Aggregates structured results
Returns them clearly to the user

The assistant acts as a trusted orchestrator, while specialized agents handle execution.

How we built it

We built AgentGate using:

A Bun backend with grammY for Telegram integration
A FastAPI-based agent network
An OpenAI-powered agent recommender
A live network visualization console built in Loveable
Real-time WebSocket event streams to track delegation and agent calls

The system supports parallel and iterative agent execution while maintaining transparency in the visual console.

Challenges we ran into

Reimagining how agents interact with the internet is ambitious. Scope creep was a constant challenge.

We explored additional problem spaces like:

Agent-to-agent billing
Payment infrastructure
Trust verification systems

Ultimately, we had to prioritize features that clearly communicated our core prototype: agent discovery and delegation.

Technically, managing WebSockets with parallel and recursive agent calls in a scalable way was also complex. Coordinating real-time updates while keeping the system modular required careful architectural decisions.

Accomplishments that we're proud of

We have a fully working proof of concept that automatically routes the personal agent to communicate to a trusted web based agent.

What we learned

Through this project, we learned:

The trade-offs between privacy-first, on-device assistants and scalable cloud-based orchestration
The importance of discovery infrastructure for agent ecosystems
How little standardization currently exists in agent-to-agent protocols
That building the “internet of agents” is as much a systems design challenge as it is an AI one

AgentGate represents a step toward a future where personal AI assistants don’t try to do everything themselves, they coordinate a network of specialized, interoperable agents.

What's next for AgentGate

Creating intranets of agents for companies to prove the economic feasibility and scalability of the concept.

Built With

ai-sdk-&-ollama-for-the-personal-ai
bun
gemini-&-openai-api-for-simulating-the-powerful-models-in-the-web.-the-visualization-is-built-with-react
python-&-uv-for-the-agentic-search-engine
typescript