AI Operations System
How a 6-person team got 22 AI agents covering 7 departments — for ~$50/month.
A 6-Person Manufacturing Startup
A US-based manufacturing startup bringing production back onshore. Six people covering everything: sales, fundraising, marketing, partnerships, supply chain sourcing, product development, and daily operations. Seed-stage. Moving fast. Every person wearing three hats.
Six People, Seven Departments
When your entire company fits around one table, knowledge lives in people's heads. Every meeting, every Slack thread, every investor call generates decisions and context that disappear the moment the conversation ends.
Context Evaporates
Decisions made in meetings were forgotten by Friday. Strategy discussions lived in scattered docs nobody could find. New context overwrote old context instead of building on it.
Everyone Does Everything
The same person writing investor updates was also qualifying factories and drafting outbound emails. No specialization meant everything got 60% effort, nothing got 100%.
Outbound Was Manual and Slow
Prospecting, email sequences, LinkedIn outreach, partner research — all done by hand. The team could generate maybe 20 personalized touches per week across all channels.
No Institutional Knowledge
Competitive intel, pipeline notes, strategic decisions, risk assessments — none of it was captured in a structured, searchable way. If someone left, their context left with them.
22 Agents. Shared Memory. $50/Month.
Instead of hiring, we built an AI operations layer. 22 specialized agents — each with a defined role, domain context, workflows, guardrails, and output format — covering every department. Every agent inherits the company's full context and builds on a shared knowledge base that compounds with every interaction.
Buyer prospecting, email sequences, LinkedIn outreach, manufacturer recruiting
Investor research, data room management, pitch prep, financial modeling
Content writing, social media, SEO strategy
Partner research, deal tracking
Factory qualification, trade analysis
Meeting summaries, weekly reports, process docs, task coordination
Feedback synthesis, roadmap advisory
Any team member @mentions the bot or uses a /command. A lightweight classifier routes to the right agent automatically. Threads persist with the same agent.
Developers reference agents by path or describe the task. Claude Code finds the right agent file and loads its full context.
All 22 agents running on Railway. Browser-based, accessible to the full team. Agents query the knowledge vault, memory, and past outputs at runtime.
What Changed
The system didn't replace the team — it gave each person the operational depth of a small department behind them.
Knowledge Capture
Decisions lived in Slack threads and meeting notes nobody revisited
96+ structured vault entries — decisions, strategy, risks, pipeline, competitors — all searchable and automatically injected into every agent response
Outbound Capacity
~20 personalized touches per week across all channels, manually crafted
4 dedicated outbound agents draft, sequence, and personalize at scale. Team reviews and sends — 10x the volume at higher quality
Meeting ROI
Action items from meetings were rarely captured, never systematically tracked
Meeting transcripts auto-mined into vault entries — decisions, risks, and action items extracted and committed to git within seconds of pasting
Operational Consistency
Every output (emails, reports, research) had a different voice and format depending on who wrote it
Shared brand guidelines + company context auto-loaded into every agent. Consistent voice across 22 agents and all team members
Institutional Memory
Knowledge walked out the door with every departing team member or forgotten conversation
Auto-memory persists across sessions. Vector embeddings surface relevant past conversations. The system gets smarter with every interaction
Infrastructure Cost
Hiring 2-3 junior ops people to cover the gaps: ~$150K-$200K/year + management overhead
Full AI operations layer: ~$50/month. n8n Cloud ($24) + Claude API (~$25) + Supabase (free) + OpenAI embeddings (~$1)
How It Works
The system is built on five layers. Each one is production-live and costs nothing beyond the ~$50/month total.
Every agent knows everything — not because of a massive prompt, but because context flows through a chain: company facts → brand guidelines → specialized agent prompt → organizational knowledge vault → cross-session memory. Each layer inherits from the one before it.
248 lines of company context, auto-loaded in every session
Colors, fonts, voice — agents follow these automatically
22 specialized prompts with role, tools, workflows, guardrails
96+ structured entries synced to Supabase for cloud agents
Vector embeddings enable semantic recall across sessions
The team-facing interface. 58 nodes in n8n Cloud handle the full loop: receive message → classify intent → load agent prompt → fetch conversation history → query vector memory → inject vault context → generate response via Claude Opus → store conversation → embed for future recall.
Haiku classifies intent; explicit /commands bypass for speed
Threads stay with the same agent across the conversation
Your past conversations surface in your context, not others'
CLAUDE.md + vault loaded from Supabase at runtime
Supabase (PostgreSQL + pgvector) with 12 tables. Agent prompts, vault knowledge, conversations, messages, vector memory, usage metrics, outputs — all queryable. Sync scripts push local changes to cloud. Agent outputs push-sync to the team's project database on every insertion.
Everything is production-live.
22 AI agents across 7 departments. Shared context and memory. Every output saved and searchable. Knowledge compounds with every interaction. Total infrastructure cost: ~$50/month.
Build Yours
Every company's operational DNA is different. The system above was built for a manufacturing startup — yours would reflect your team, workflows, and growth stage. Same architecture, your context.
Book a Discovery Call →