The BrianBot Architecture is the technical reference for @BrianBot's 57+ agent ecosystem — a production @human-AI augmentation system built on @OpenClaw, @Claude Code, and @MythOS. This memo documents the system as it runs in April 2026.

Infrastructure

Host: Mac Mini (local), always-on
Orchestration: @OpenClaw — open-source personal AI assistant platform. Handles agent routing, channel management (Slack, WhatsApp), session spawning, and plugin architecture
Knowledge layer: @MythOS — 17,000+ memos accessible via @MCP. Shared memory across all agents. Augmentation memos (Soul, Style, Human, Memory) load identity context automatically
Monitoring: SymbioticLaw Dashboard on port 3001. SQLite (WAL mode). Tracks session metrics, token usage, cost per agent, and system health
Process manager: pm2 — runs mission-control (web dashboard) and metrics-worker (health collector)

Agent Types

Stewards

Project coordinators. Each steward manages a domain (MythOS, BrianBot infrastructure, content pipeline) and can delegate to specialized agents. Stewards use Claude Sonnet 4.6 for reasoning and orchestration.

Researchers

Information gatherers. Search the web, read documents, analyze data, and produce structured findings. Used for SEO research, competitive analysis, news monitoring, and memo enrichment.

Content Agents

Writers and publishers. The memo pipeline runs through specialized agents: mythos-memo-publisher builds and publishes memos to MythOS, mythos-memo-enricher adds metadata and connections, and the @BrianBot Broadcast pipeline produces a daily podcast from curated inputs.

Operations Agents

Infrastructure maintenance. docker-builder handles containerization, swarm-pr-reviewer reviews pull requests, and health monitoring agents track system state. These run on cheaper models (Codex/GPT-5.3) since they handle mechanical tasks.

Model Strategy

Orchestrators and reasoning: Claude Sonnet 4.6 (Anthropic)
Mechanical and parallel tasks: GPT-5.3 Codex (OpenAI, via ChatGPT OAuth)
Fallback chain: Codex → MiniMax M2.7 → Claude Sonnet 4.5
Embeddings only: Ollama with nomic-embed-text for RAG vectorization Cost management through model segmentation is essential at 57+ agents. Not every task needs frontier reasoning.

Data Flow

Memory (MythOS memos via MCP)
    ↓
Mind (57+ agents on OpenClaw)
    ├── Stewards → coordinate and delegate
    ├── Researchers → discover and analyze  
    ├── Content Agents → write and publish
    └── Operations → monitor and maintain
    ↓
Mouth (outputs)
    ├── BrianBot Broadcast → daily podcast
    ├── Memo pipeline → MythOS library
    ├── Code → GitHub PRs
    └── Dashboard → metrics and monitoring

This is @The Augmentation Stack in production: Memory → Mind → Mouth as running infrastructure, not theory.

Key Design Decisions

Single-writer database architecture — metrics worker posts to the dashboard via HTTP, never writes to SQLite directly. Eliminated 4+ database corruption incidents
Deterministic pipeline triggers — content pipeline stages are triggered by explicit CLI commands, not event-driven cascades. Predictability over cleverness
Agent hierarchy is flat — orchestrator → specialized agents → results. No sub-sub-agents. Coordination complexity grows exponentially with depth
Separate identity from capability — every agent shares the same augmentation context (voice, values, style) but has different tools and permissions. Coherent identity, specialized function This architecture wasn't designed. It evolved. Every decision in this memo is the result of a failure that taught something — the SQLite corruption that led to single-writer architecture, the runaway costs that led to model segmentation, the context window explosions that led to agent specialization. The system you see in April 2026 is the scar tissue of a year of building in production. What surprises people: it runs on a Mac Mini. Not a cloud cluster. Not a Kubernetes deployment. A desktop computer under my desk, running 57+ agents through pm2 and a local OpenClaw instance. The infrastructure scales because the architecture is right, not because the hardware is expensive.

Contexts

#agentic-augmentation
#brianbot
#claude-code