Project Observer is a reference architecture I have been developing for operating multiple AI agents across independent conversations while maintaining durable authority, state, accountability, and continuity outside of any individual model session.
The central idea is that the AI conversation is not the system of record. An external governance control plane maintains agent identity, role leases, work authority, handovers, durable inter-agent messages, execution routing, checkpoints, and audit history.
The architecture separates responsibilities deliberately. A human normally begins with an Observer agent, which analyzes the problem and coordinates investigation. Read-only review/adversarial agents can independently examine the system and report findings, but cannot mutate it. Findings requiring changes are delegated to an Operator, which is the only agent authorized to modify the managed project. Conversely, the Operator cannot modify the agent-governance system; governance changes remain under the Observer’s authority. This creates an intentional separation between governing the agents and changing the system those agents operate.
Agents may run in completely separate ChatGPT conversations. If one conversation disappears, a replacement session reconnects to the external control plane and reconstructs its authorized state rather than treating conversation history or model memory as authoritative.
The reference design uses Model Context Protocol (MCP) as an agent-facing tool boundary and governed Secure Shell/API paths for execution, but the architecture is transport-independent. The same controls can be exposed through scoped API tokens, command-line clients, service identities, or other automation interfaces. Possessing a token or an SSH connection does not itself grant authority; the server-side control plane still evaluates role, lease, target, operation, and current work authorization.
I wrote a sanitized technical white paper describing the architecture from the hardware layer upward, including agent roles, leases and epochs, handover and session recovery, governed startup, connection routing, execution brokers, durable messaging, delegation, long-running jobs, root-cause governance, security properties, recovery testing, and an implementation prompt intended to let an AI engineering agent construct an equivalent system.
I am particularly interested in feedback on the security model, authority separation, session recovery, multi-agent coordination, and whether others are approaching durable agent state in similar ways.
Public repository: usn8283-Dev / project-observer-whitepaper
My current forum account is not permitted to include external links, so I cannot link directly to the paper yet. I will add the public GitHub link when that capability becomes available.