Case Study 05
Agent Team
Autonomous Platform
Product
Ivyna — AI-managed tech agency
AI Runtime
Claude Agent SDK (Sonnet)
Scope
Full autonomous business ops
Specialized AI agents
Custom MCP tool servers
Autonomous tick cycle
Continuous operation
The Concept
What if an entire agency was run by AI agents?
Agent Team is an autonomous multi-agent platform built on the Claude Agent SDK. It runs a team of 7 specialized AI agents that collaborate to handle business operations — content creation, lead processing, CRM management, outreach, and email — all orchestrated through an event-driven lifecycle engine.
The system operates continuously with minimal human oversight. A scheduler ticks every 5 minutes, checking agent agendas and creating proposals automatically. A policy engine decides what can be auto-approved and what needs human sign-off. Safety guardrails cap daily proposals, concurrent missions, and execution timeouts.
This isn't a chatbot with tools. It's a self-directing organization where agents research, create, review, publish, and follow up — autonomously.
The Team
Seven agents. Each a specialist.
Orchestrator
Decomposes tasks, routes to specialists, manages mission lifecycle
Research
Web research, lead qualification, competitor analysis, market intelligence
Engineering
Code generation, debugging, Git operations, technical implementation
Content
Blog posts, social media copy, LinkedIn carousels, thought leadership
Marketing
Campaigns, SEO strategy, audience analysis, growth planning
Documentation
READMEs, API docs, architecture docs, technical writing
QA
Reviews all outputs — accuracy, grammar, quality scoring on 100-point rubric
Autonomous Lifecycle
Five components that make it self-directing.
Lifecycle Engine
Scheduler
Ticks every 5 min, checks agendas, creates proposals
Proposals
Work requests from users, triggers, scheduler, or agents
Missions
Multi-step execution plans with QA at each step
Triggers
Event chains — Research → Content → Marketing
Policy
Auto-approval rules, budgets, safety gates
Kill switch
Instant halt
50/day cap
Proposal limit
3 concurrent
Mission max
30 min timeout
Auto-cleanup
Business Capabilities
Six domains. Fully autonomous.
Content Pipeline
Draft → structural review → duplicate detection → AI marketing review (100-point rubric) → approve/revise/reject → publish → post to LinkedIn and Twitter
CRM
Contact management, deal pipeline (discovery → proposal → negotiation → won/lost), AI-generated summaries for every interaction
Lead Processing
Public contact form → auto-create contact → Research agent qualifies → LinkedIn outreach sequence fires automatically
Email Operations
Resend API for outbound, Outlook Graph for inbox monitoring, AI-drafted responses, full thread tracking
Pricing & Quotes
Service pricing rules, auto-generated quotes from deal context, negotiation log for every interaction
Outreach Sequences
Multi-step LinkedIn automation — connection request → intro DM → follow-ups — with automatic stage advancement
Tech Stack
Claude Agent SDK
AI runtime (Sonnet)
FastAPI
Backend API (port 8001)
Supabase
PostgreSQL + pgvector
Celery + Redis
Task queue
Playwright
LinkedIn/Twitter automation
Resend + Graph
Email (send + inbox)
10+ MCP Servers
Custom tools (port 9100)
React 19 + Vite
Admin frontend
Phaser.js
Pixel Office visualization
Langfuse
Observability (Docker)
Tailwind CSS 4
UI framework
WebSocket
Streaming chat UI
The Vision
Agent Team proves that a multi-agent system can run real business operations — not just answer questions, but research leads, write content, review quality, publish to social media, process inbound leads, and follow up via email — all while maintaining human oversight through a policy engine.
The Pixel Office visualization (built in Phaser.js) makes the invisible visible: you can watch agents move between desks, collaborate on missions, and report results in real time. It turns autonomous AI from a black box into something a team can observe and trust.
Autonomous agents
Business domains
Continuous operation
Next project