Agentikas Blog is an AI-first, open-source blogging platform — free for everyone, with every post visible to AI agents by default. Read articles from our lab and the community, or create your own blog in 2 minutes.
Day 4 was meant to test the 5-agent pipeline against real Claude. It ended with 4 hours fighting Anthropic billing, a clear architectural lesson, and an Ollama + Qwen 2.5 Coder 32B integration that's worth more than the original plan.
Day 2 left the agents deciding architecture without real repo context. Day 3 introduces MCP (Model Context Protocol) as a complement to A2A: a Codebase MCP server that exposes the code, and a Project Mapper agent that consumes it via the Anthropic SDK's native MCP integration.
The initial plan called the second agent "Auditor". Within hours we changed the name to "Project Explorer" / Tech Lead. It's not cosmetic — it's the difference between a system that finds problems and one that designs solutions. This is the lesson of Day 2.
We built the functional skeleton of an agent network communicating via A2A from scratch. Three services, a conversational Orchestrator, a Notifier, and an Investigator with a real LLM. Real bugs included.
An agentic software development architecture where human experience gets codified into skills, schemas, and names. It works — but without a senior behind it, it's the perfect recipe for technical debt at AI speed. The MVP, the 5 agents already running, the 2 yet to come, and why the tool doesn't replace judgment.