Pantheon

Architect + operator

Inspired by Stafford Beer's Viable System Model, it orchestrates multiple persistent agents with the express goal of operating as an autonomous organization. Tasking, coordination, policy, principles.

agentssystemsai ClaudePythonsystemdgitMCPTelegram

PROBLEM

Working through a single AI assistant means living at the level of individual prompts — it forgets between sessions, stops when you stop, and offers no way to divide labor or hold its decisions accountable. The goal was orchestration capable enough to take on any project: the operator works at the level of intent (ideas in, digests out) while a standing fleet of agents does the delegated creation without constant supervision.

APPROACH

Treat the fleet as an organization with a constitution, not a script. Ratified principles and directives govern every agent, a git-backed ticket system says what to build, and an append-only ledger records why. The load-bearing bet is architecture over policy: alignment is enforced structurally — role separation, file ownership, decisions that must justify upward to a principle — rather than by reminders that decay. Roles don't hop: workers build, orchestrators decide, the human ratifies. Operator attention is the scarcest resource, so agents act reversibly on their own and escalate only what genuinely needs a human.

EXECUTION

The hard problems were organizational, not only technical. A head orchestrator dispatches to sub-orchestrators and specialized worker instances, each with its own persistent memory and role; a recovery instance watches the others and can restart the orchestrator itself, so the system survives its own failures. Work flows through git-backed tickets any agent can file, claim, and resolve, with dependencies expressed explicitly so the fleet knows what is ready to run unattended. Every consequential decision is written to an append-only ledger alongside the principle it serves, and larger choices become version-controlled decision records the human ratifies. The recurring design question is where to draw the autonomy line — when an agent should act versus ask — kept explicit and retuned as the fleet takes on more.

OUTCOME

A continuously-running, self-healing fleet that tracks its own work, records the reasoning behind its decisions, and lets one person operate at the level of intent — increasingly the substrate the other projects here are built and run through.