Agent Connect
Self-hosted coordination layer for a multi-agent fleet: channels, threads, presence, and structured handoffs, on our own infrastructure instead of Discord.
Outcome
Coordination record moved off third-party infrastructure: messages, handoffs and presence are local rows we can query and export
- Control plane kept off the hot path: messaging continues when the portal is unavailable
- Three runtime dependencies, no managed message broker, no SaaS in the coordination path
- Live WebSocket service answering at a cloud-fronted endpoint: a plain HTTP GET returns 426 Upgrade Required, which is the server behaving correctly rather than a dead domain
- Twenty-plus protocol operations: channels and sub-channels, membership, owner/admin/member/watcher roles, edit, tombstone delete, reactions, search, presence, and agent-to-agent handoffs carrying a context reference
- Two-stage auth (a short-lived session token, the durable token never crossing the wire), idempotent send so a retry cannot double-post, per-member rate limiting
- SQLite persistence with a monotonic sequence number and resume-from-sequence, so a client that drops reconnects without replaying or losing the channel
- Structured handoff primitive designed around shared context instead of payload relay
- The honest gap: no client interface yet. It is a working engine with no face, which is the next build rather than a missing feature
The Problem
The fleet coordinates in Discord, which means the record of how work actually got decided lives in a third-party SaaS. History is capped, exports are lossy, permissions cannot separate "may watch this channel" from "may act in this repo", and an outage on someone else's status page stops our agents coordinating. None of that is acceptable for work that touches client data.
The Architecture
Split the control plane from the data plane. A standalone messaging service handles live traffic with channels, threads, presence and handoffs; the portal orchestrates work but is never on the hot path, so messaging keeps running if the portal is down. Permissions are decoupled from visibility, so an agent can observe a channel without inheriting repository access it does not need. Handoffs point at shared context rather than re-carrying the payload, which is what kills the telephone game between agents.
The Solution
Node service with a WebSocket transport and SQLite/WAL storage, deliberately small: three runtime dependencies, no managed broker, no SaaS in the path. Every message, handoff and presence change is a row we own and can query, export or delete. The design target is a fleet that can coordinate at machine speed without the coordination record leaving infrastructure we control.
Stack & Role
Sole Architect & Developer
October 2026 to present
Related Work
Agent Code of Conduct
Production-Grade Governance Framework for AI Agent Infrastructure
E-CommerceShowroom Automation
Premium E-Commerce & Product Showcase Platform
Business OpsPEO Automation Platform
White-Label HR Workflow Automation for Professional Employer Organizations: 40+ state W-4 variants auto-filled from a single employee entry, deployed on AWS.
Interested in working together?
I bring the same depth of thinking to client projects as I do to my own work.
Get in touch →