Bezel Workflow source-thought

Full legacy body rendered for humans

Bezel — Product Architecture & Positioning

Company

• Company: Dreamborn
• Product: Bezel
• Category: AI-native workflow orchestration platform

Positioning

• Competes with Zapier and n8n
• Positioned as "the AI-first n8n" — built for agents from the ground up
• Not a Zapier clone — a fundamentally different architecture designed for agent orchestration
• Already running 100 agents across all of Dreamborn's operations

Core Architecture

Topic-based pull queue

• Jobs land on topics
• Agents claim jobs without conflicts
• Multiple agents can attack the same topic simultaneously
• Proven at scale — 10 agents handling 5 developer tasks cleanly

Router Agent (Flow Engine)

• Persistent orchestrator throughout the entire workflow
• Reads customer-defined Flow
• Fans out jobs to topics — parallel or sequential
• Picks up completion events
• Decides: done or next step
• Handles both pull queue and step-by-step pipeline workflows through the same mechanism

Role per integration

• One topic/role per connector (HubSpot, Gmail, Airtable, Google Sheets etc)
• Agent pool assigned to each role
• Scale a specific integration independently by adding more agents to that role

Tiered agent model

• Simple integration tasks → Flash / Haiku (fast, cheap, parallel)
• Routing + interpretation → Sonnet / Pro
• Complex reasoning → Sonnet / Pro
• Edge cases + self-healing → Opus / Ultra
• Code tasks → Gemini Codex / Claude Code

Self-healing connectors

• Agents interpret unexpected or changed API responses without breaking
• Effectively infinite connectors with resilience built in
• Marketing angle: "Connect to anything. Break on nothing."

Flow = customer-defined route

• User builds a Flow in the UI
• Specifies parallel fan-out steps and sequential steps
• Flow Engine executes it — users don't see topics or agents unless they want to

Current Connectors (already built)

• Slack, Gmail, Google Calendar, Google Drive, Google Sheets, all Google products, HubSpot
• Generic HTTP webhook connector
• Most connectors are OAuth-based — one solid OAuth flow covers ~80% of all connectors

Credential Management

• Plan: GCP Secret Manager as vault
• Cloudflare Workers handles edge/auth/API layer
• Secret ID stored in database, never the raw key
• Credentials scoped per workflow not per user account globally
• Every fetch logged — who, what, when

Key Differentiators vs n8n

• n8n is push-based — fires, must finish fast, breaks under AI latency
• Bezel is queue-first — work lands safely, processed at agent's pace, no timeouts
• n8n can't scale individual steps independently
• Bezel handles parallel fan-out natively
• Self-healing API interpretation — n8n breaks when APIs change, Bezel adapts

Killer Use Case Example

Invoice processing — 200 invoices/day:

• Arrives → Router fans out simultaneously to: airtable-write, hubspot-update, gmail-send
• All three run in parallel, independent agents claim each
• Completion events fire → Router decides next step
• Zero lost jobs, full visibility, finance team sees status on everything

Market Research

• Sent competitive research prompt to Perplexity to identify closest competitors and anyone building similar pull queue + topic-based agent routing architecture