Engineering patterns, build logs, and field notes from running a production AI agent network.
Vertex AI Pipelines, custom BART models on GPU, three separate AI services per review. I ripped it all out. One Gemini Flash call, Cloud Workflows, and Dataform replaced everything at 98% less cost.
Six mistakes every team makes building agent systems, and the architectural patterns that fix them.
The MCP documentation shows you how to build a toy server. Running one in production means solving transport reliability, tool schema design, auth patterns, and failure modes that no tutorial covers.
Seven months deep in Claude's ecosystem. Desktop, CLI, mobile, Projects, MCP servers, multi-agent fleets. What works, what burned me, and why I can't stop ordering.
The smartest tools in my stack were bottlenecked by the only component that needed coffee and sleep. So I built CacheBash, an open-source MCP server with 34 tools that lets your AI sessions coordinate without you in the middle.
Each tool is useful. The Heartbeat saves me from checking tmux. The Task Queue saves me from copy-paste handoffs. But when you connect all six, something different happens.
Five AI agents running in parallel. One git checkout and the whole thing collapses. Real lessons about isolation, session death, and fleet management from running multi-agent AI in production.
Design patterns for bounded, vendor-neutral multi-agent systems in production. Most teams think orchestration equals prompt chaining. That's plumbing. The hard problems are identity, mortality, trust, and economics.
CacheBash started as an internal tool for coordinating AI sessions. Three months later, it's an open-source MCP server with 34 tools, a mobile app, and a roadmap toward commercial infrastructure.
Your AI agent has shell access and API keys. A fail-closed compliance layer stops every ambiguous action. A fail-open one logs everything and lets the work continue. We chose fail-open.
You open three terminal tabs. Claude Code in one, Cursor in another, a third deploying to staging. None of them know the other two exist. I built CacheBash to fix that.
Most planning tools treat specs as a gate — pass/fail. But a 10-step plan where step 7 is 'figure out the database schema' isn't much better than no plan at all. specfirst adds a gradient.