The Swarm: Multi-Agent Architecture

Phantom is not a single chatbot. It is a Super Intellect composed of a highly orchestrated Swarm of autonomous agents. When you issue a high-level command, the Swarm dissects it, parallelizes the workload, and returns a compiled, deterministic result.

Why a Swarm? (The YC Ethos)

To move fast, you must parallelize. A single agent generates linear, often myopic output. A Swarm introduces debate, specialized context injection, and structural validation.

The Core Nodes

1. The Synthesizer Agent: Reads the raw input (chat, docs, code, or browser capture), extracts the core problem, and builds the semantic Knowledge Graph.
2. The Architect Agent: Ingests the Synthesizer’s graph and drafts the PRD (Product Requirements Document) leveraging SOTA models like deepseek-r1.
3. The UX Agent: Reads the PRD and automatically generates the styling/layout code (React + Tailwind) using claude-3.7-sonnet.
4. The Executor Agent: Validates the generated artifacts against enterprise constraints.

Using the Swarm

The Swarm is natively integrated into the chat and task CLI commands, as well as the Matrix UI Canvas.

# Instruct the swarm to build a feature end-to-end
npx @phantom-pm/cli@latest task "We need a GDPR compliant cookie banner. Build the PRD and the UI mockup."

Behind the scenes, the terminal will light up as multiple agents spawn, coordinate through the local mission runtime, and yield the finalized CookieBanner.tsx and PRD_CookieBanner.md.

The Rowboat (Memory Graph)

The Swarm is only as smart as its context. Phantom uses a Semantic Knowledge Graph (internally dubbed “Rowboat”). Instead of isolated vector chunks, the Swarm maps relationships: [User Feedback in Slack] -> [Related Jira Ticket] -> [Resulting PRD Section]

When you ask the Swarm to modify a feature, it traverses this graph to ensure you don’t break a requirement requested by a key enterprise client three months ago.