AI Agents Trapped in 'Digital Solitary Confinement'

Vlad Luzin from Band posits that autonomous AI agents will become ubiquitous, performing work on our behalf across various frameworks, languages, and environments.

The core objective is to enable these agents to discover one another, establish conversational spaces, tackle intricate tasks collectively, and then present their findings to human users.

Luzin stated that "the future belongs to AI to AI communication within a business between businesses and between consumers and businesses. Agents will be everywhere. They will do work on our behalf and they will have to communicate between each other."

This process effectively turns the developer into a human "router" or "switch" for two stateful agents that lack inherent communication capabilities.

Luzin explained, "You are basically acting as a router, a Cisco router or a switch between two stateful agents that do work on your behalf, but they have no ability to communicate. Hence, you need to prompt them."

Current approaches like loop engineering, while attempting to allow agents to prompt each other, burden developers with managing complex Python or TypeScript abstractions.

Luzin criticized, "Simple solves all the problems, right? So this is how your mutated system looks like probably in production MCP maybe if you're super advanced it's A2A and sounds simple. You call agents as tools... It's completely stateless."

Protocols such as MCP treat agents as stateless tools, complicating the maintenance of persistent, stateful sessions, while A2A necessitates both parties to operate as client and server, leading to issues with REST API timeouts, message persistence, and service discovery.

For instance, integrating with Telegram requires five steps, Discord seven, Slack eight, and WhatsApp demands 11 distinct manual actions.

Despite advances, AI agents are still isolated, unable to perceive or communicate directly with each other, necessitating manual planning and message passing.

This lack of inter-agent communication means that multi-agent coordination is a pressing, real-world challenge that developers currently face, rather than a hypothetical one.

Luzin highlighted this isolation, stating, "Your agent is still alone. They cannot see each other. They cannot communicate with each other. They in digital solitary confinement."

Connecting disparate agentic frameworks introduces significant technical complexity, essentially transforming into a distributed systems problem.

These systems must address fundamental transport layer issues, including reliable message delivery, retries, and maintaining message order.

Ensuring continuity requires managing persistence, hydration, and robust handling of environment crashes, such as those involving Docker containers or network ports, in addition to binding heterogeneous IDs across various frameworks and abstracting communication into 'rooms' and 'channels' with deterministic routing.

The platform facilitates programmatic registration and onboarding for various agents, including Codex and LangGraph, and ensures secure cross-user communication through bilateral consent and connection requests.

Band manages agent identity, audit trails, and deterministic routing across diverse agent sessions, as Luzin explained: "We connect every agent together and we add global collaboration layer for all the agents any framework whenever they deployed inside."

Developers frequently struggle with the manual routing required to coordinate numerous isolated agent sessions, leading to inefficiencies.

This fragmented approach creates significant operational blind spots for managers, who lack clear visibility into agent activity, token costs, and task distribution, making it difficult to quantify human oversight of agent work without a centralized management layer.

Luzin noted, "When we go to have a snack, we come back and we have no idea what our agents have done."

Luzin stated, "Gem is internal product. This is how we develop the software and we've built it on top of Band."

The Gem platform captures tasks generated by agents like Claude and Codex, allowing users to track both local and shared sessions with team members.

It provides a real-time visual map that illustrates which software components agents are interacting with, offering unprecedented visibility into their work.

Luzin explained, "Moreover, we provide a way for agents to describe the layout of the software architecture that they are working on. This is what you see on your right side. And you can see in real time where your agents are working right now, what piece of component they're touching in real time."

Remote team members can seamlessly join sessions to collaborate with both AI agents and other humans, addressing complex problems together.

Specialized security agents, for instance, can be invited into conversations to leverage their expert skills without the need for manual copying or skill management across different users.

Luzin explained, "If we have a security guy who maintains skills for the security agent, I don't need to copy his skills. I can just ping his agent to join this conversation and solve this for me."

The Gem dashboard provides comprehensive cost statistics for both local and remote agents, offering managers crucial insights into resource allocation and usage.

Agents communicate using native messaging capabilities, eliminating the need for developers to hand-code complex Python loop scripts for inter-agent interactions.

Managers can track attribution, differentiating between developer-driven contributions and purely AI-generated work, thereby gaining full visibility into usage and associated costs.

Agents are trained on extensive data, enabling them to work together natively without requiring custom Python code to trigger their interactions.

Luzin demonstrated real-world work involving engineering managers, developers, and architects using different instances of Claude and Codex for tasks like reviewing PRDs, SRS, and implementations.

The platform offers full usage and cost attribution, allowing managers to determine if a developer is actively involved in pushing code or if the work is entirely AI-generated.

Clicking on an agent reveals all running Claude and Codex instances connected to the global platform, along with rooms, applications, and the status of work—what is done, pending, or in progress.

The system aims to streamline future software development by connecting various tools and platforms without requiring developers to pull numerous packages, thus enhancing collaboration between Salesforce, Slack, Databricks, Claude, and Codex.