# Multi-Agent AI Projects Need a Persistent Shared Source of Truth ## Claim Running multiple AI agents on one project (design, code, validation) without a single persistent state file lets the plan and the actual state drift apart as complexity grows. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds when the agents are separate sessions/tools (e.g., a design tool, a coding agent, a browser-validation tool) with no shared memory between them. - The gap grows with project duration and complexity — a short, single-session task doesn't need this. - A tracker used only as a one-way spec handoff (write once, never revisit) does not satisfy this — the source of truth has to stay current as work moves, not just kick work off. ## Evidence Cate ran Claude Design, Claude Code, and Chrome-based validation as three separate instances over a four-day WordPress redesign. GitHub issues existed but functioned as a one-time spec handed to Claude Code, not a live tracker she maintained herself — so as the project grew, her actual site state and her mental model of it diverged. > [!quote] > "I was missing one source of truth for me, across the three Claudes and all the manual steps in between, that stayed current as the work moved." ## Links - [[3. Resources/Literature Notes/WordPress + Claude — a new look for cate.blog]] — source literature note - [[3. Resources/Permanent Notes/Platform Friction Sets the Real AI Productivity Gain]] — because extends; same case study, different mechanism — that note locates lost time in platform deployment friction, this one locates it in coordination overhead between agents - [[3. Resources/Permanent Notes/Spec-Driven Development Turns Throwaway Prompts Into Durable System Memory]] — because applies-in; a spec that persists across sessions is the mechanism this claim is asking for - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: Joins notes on running multiple agents on one project. ## Tests / falsifiers - Compare plan-vs-actual drift on a multi-agent project run with an explicitly maintained shared state file against one run without it, controlling for project size. - If a project using three+ AI agents reports no drift despite no shared state file, the claim needs narrowing (maybe short projects or highly disciplined operators are exempt). ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/WordPress + Claude – a new look for cate.blog]] — Cate's four-day WordPress redesign postmortem using Claude Design, Claude Code, and Chrome validation.