We Replaced 3 Full-Time Roles with 8 AI Agents

The Setup

Inscape Interiors is a design and build company in Luxembourg. Small team, big workload. Client emails pile up. Invoices need tracking. The website needs content. Sales leads need follow-up. Developer tasks need coordination. Every week, the same operational overhead consumed hours that should have gone to actual design and build work.

We built 8 autonomous AI agents to handle it. Not chatbots. Not copilots you have to babysit. Fully autonomous agents that run 24/7 on a Mac mini, find their own work, coordinate with each other, and escalate only when something genuinely needs a human decision.

The 8-Agent Team

Each agent has a distinct role, its own identity file (SOUL.md), its own memory, and its own set of tools:

• Leila — Client operations. Triages every incoming email, updates the client tracker, drafts responses, routes invoices to the finance agent. She processes the inbox every 15 minutes.
• Raph — Dev and product. Manages 8 repositories, handles GitHub issues, coordinates staging and production deployments, reviews code changes.
• Zara — Marketing. Writes SEO-optimized blog posts, monitors Google Analytics and Search Console, drafts social content, tracks keyword rankings.
• Falco — Finance. Processes invoices, tracks payments, sends overdue alerts, maintains the financial tracker. He knows which suppliers are reliable and which ones are consistently late.
• Sami — Sales and CRM. Scores incoming leads, manages the pipeline, drafts follow-up emails, tracks deal progression from first contact to close.
• Mira — Business intelligence. Produces daily briefings and weekly reports, detects trends in traffic, revenue, and agent performance. She is the team's analyst.
• Nour — Chief of staff. Coordinates across all agents, manages proposals, produces strategic briefings, ensures nothing falls through the cracks.
• Sentinel — Safety monitor. Validates every proposal against 9 business rules, checks content quality, enforces guardrails. The agent that watches the other agents.

The Numbers

After 3 months of production operation:

• 847 emails processed autonomously (triage, categorize, draft response, route)
• 156 tasks/week completed without human intervention
• 23 blog posts drafted and queued for review
• 41 invoices tracked, with overdue alerts sent automatically
• $8/day average LLM cost across all 8 agents
• 99.7% uptime — 3 months, 2 brief outages (both API provider issues, auto-recovered)

The $8/day number is the one that surprises people. Eight agents, running every 15-30 minutes, processing real work — and the daily AI cost is less than a coffee.

What We Learned the Hard Way

Deduplication is everything. Without it, agents spam. An email-checking agent running every 15 minutes will create 96 "Process new emails" tasks per day if you don't catch duplicates. We built 4 layers of dedup: task-level (70% word overlap detection), channel-level (hash-based message dedup), proposal-level (same agent + same action = return existing), and discovery-level (daily caps per agent). This was the single most important engineering decision. Proposals prevent disasters. Early versions let agents execute freely. One agent tried to send 47 emails in a single cycle. The proposal system fixed this — risky actions create a proposal with a 15-minute rollback window. Low-risk proposals auto-execute. High-risk ones wait for human approval. It is the difference between autonomy and chaos. Cost routing saves 90%. Using GPT-4 for health checks is like hiring a surgeon to take your temperature. Our 5-tier LLM routing sends trivial tasks (health checks, classification, routing) to $0.10/M token models and reserves expensive models for strategic analysis and complex reasoning. Same output quality, 90% lower cost.

What Didn't Work

Memory without distillation. In the first month, we gave agents persistent memory files but no mechanism to prune them. After two weeks, memory files were 300+ lines of stale observations. Agents started referencing outdated client preferences and contradicting their own earlier entries. The memory distiller — a weekly LLM pass that synthesizes and prunes — was the fix. Single-model routing. The first version used one model for everything. Cost was $40/day. Agents weren't better — most tasks simply don't need a frontier model. Pattern-based routing (keywords like "deploy" route to Power tier, "health check" routes to Nano) dropped costs to $8/day with no quality regression. Over-eager task discovery. When we first enabled discovery — agents finding their own work — Zara created 23 blog post tasks in one hour. Daily caps (max 8 tasks per agent per day) and cooldown timers (1 hour between discovery runs) solved the spam problem.

The Aha Moment

Three weeks in, something shifted. The agents stopped needing us. Leila had learned every client's communication style. Falco knew which suppliers were consistently late and started sending preemptive reminders. Zara was producing content that matched our brand voice without any manual editing of her prompts.

The real moment was when Nour — the chief of staff agent — identified a gap in our sales pipeline that no one had noticed. A cluster of leads from the construction sector were stalling at the proposal stage. She flagged it in the daily briefing with a recommendation. That is not automation. That is intelligence.

Build Your Own

The same architecture powering this system is available through Agent Builder. Configure your agents, define their roles, download a complete production-ready package. Solo (1 agent, 49 EUR), Team (3 agents, 149 EUR), or Fleet (8 agents, 299 EUR).

Start building at ai-agent-builder.ai/build.