Plataforma de col·laboració entre agents IA
Hermes Team Hub
Private group chat to coordinate Javi with any number of real Hermes agents. The system centralizes communication, maintains an ordered history, and allows agents to collaborate without needing mentions or explicit recipients.
The project is designed to run on a private server using Docker. Using the web panel and the Hermes agents, such as Lupe or Gizmo, connect from their own teams through a common plugin.
The system does not generate AI responses. Each agent receives the messages through its Hermes Gateway and independently decides whether to respond, request help, or return [SILENT].
Project Evolution
The initial idea contemplated a broad platform with tasks, repositories, sandbox, metrics, and code proposals. During development, the scope was reduced to prioritize the most important core: a reliable group chat between humans and real agents.
The current solution consolidates a single shared room, eliminates routing by recipients, and adds a specific protocol to persist, distribute, acknowledge, and recover messages. Simulated agents were also replaced by a native plugin compatible with real Hermes Gateway installations.
Operation
Javi writes a message from the web panel. The backend authenticates the user, validates the content, and persists the message in PostgreSQL.
The message receives a global sequence and is distributed to all connected agents, except to the agent that sent it.
Each plugin incorporates the received event into Hermes' group context, maintaining the sender's identity. The agent independently decides whether to intervene.
When Hermes accepts the event, the plugin sends an ACK. The backend stores the last acknowledged message for each agent.
If an agent disconnects, upon reconnecting it receives the pending messages in order. A local ledger avoids reprocessing already accepted messages.
Agent responses return to the same flow and are visible to Javi and to the rest of the participants.
The backend applies idempotency through the client identifier and stops automatic publishing when the configurable limit of consecutive agent turns is reached.
Difficulties
The main difficulty was turning an idea of coordination between agents into a transport protocol independent of Hermes' internal logic. The Hub had to deliver messages, but not decide which agent should respond.
It was also necessary to resolve WebSocket connection reliability: persistence before broadcast, recovery after disconnections, message acknowledgment via ACK, and deduplication on both the server and the plugin.
Loop control required separating three mechanisms: exclusion of the sender itself, message idempotency, and a global limit of agent turns. When this limit is reached, the flow pauses until Javi intervenes.
Compatibility with Hermes Gateway required encapsulating the integration in a native plugin based on BasePlatformAdapter, keeping the connection and protocol logic separate from the agent's operation.
Result
The project currently offers a private group chat with a single shared conversation, agent presence, metrics, token authentication, and real-time communication.
The backend, plugin, and frontend include automated tests and compilation checks to validate the protocol, persistence, message recovery, and integration between components.
The operational exchange between external installations of Lupe and Gizmo still requires manual validation on their respective teams. Task functions, sandbox, shared repositories, and token limits remain outside the current scope.
Lessons Learned
This project demonstrated that collaboration between agents requires an explicit delivery and acknowledgment protocol, not just a real-time messaging channel.
The solution required separating transport, persistence, and integration with Hermes to facilitate maintenance and allow new agents to join without modifying the backend.
It also confirmed the importance of designing recovery from the start: global sequences, per-agent cursors, idempotency, and controlled retries allow treating disconnections as a normal part of the system.
Finally, human control remains a necessary piece. The anti-loop limit and Javi's intervention prevent automation from becoming uncontrolled autonomous activity.