AI agents, automation and infrastructure.
End-to-end architecture for agent systems, explained from real implementation with the terminal open. Google ADK, A2A, Claude Code, MCP and everything it takes for an agent to survive production.
End-to-end architecture for agent systems, explained from real implementation with the terminal open. Google ADK, A2A, Claude Code, MCP and everything it takes for an agent to survive production.
New course · in Spanish · 7 levels · 2 h 44 min
From a real repository to production: seven levels, evals and a capstone where an agent takes an issue and the pipeline promotes it to production with nobody at the keyboard. Guide and video in Spanish.
The term was born from a 12-word tweet; the pattern is years old. The 5 waves, and why it doesn't replace the Loop or the Harness.
The term is marketing, but the change underneath is real. The genealogy, the verification that changes everything, and 7 agents looping in production.
How Uber connects 1,500 agents to 10,000 services with one protocol, Soria Parra's thesis and the roadmap.
Why Claude, GPT and Gemini no longer matter. Agent = Model + Harness, the discipline adopted by OpenAI, Anthropic and Karpathy.
Agents that execute skills and discover each other: Python, ADK, Claude Code with SKILL.md and A2A in one system.
Two rival AIs, Claude and Codex, resolved a real production incident. Caged, and with a human approving.
See the case →
The open-source AI brain from Y Combinator's CEO, and the sovereign harness I built on top: one memory shared by Claude, ChatGPT and Gemini.
See the case →
The Architect reasons the stack and SequentialAgent coordinates the 7 agents that build it, via A2A artifacts, to generate a complete IDP.
See the case → Docs →
7 agents build a skill factory: 4 instances debate and build, 3 subagents execute scoped tasks. One prompt, the skill published.
See the case → Docs →
7 ADK agents as independent A2A servers. Each with its own URL, its own container, and scales to zero. No shared filesystem.
See the case → Docs →What they actually are, and the fine print the announcement leaves out.
Anthropic announced Claude will mark everything it generates. I tested it with real evidence: what C2PA is, why it disappears on its own, and how it compares to SynthID.
Not one product, but four on the same agent harness. Free models, the CLI bridge and the clause that bans Claude Code.
The AI employee that lives in your Slack. Under the hood it's Claude Code, and it has one catch: where its memory lives.
The model the US government shut down, the fine print of its return, and the experiment proving the silent fallback to Opus 4.8.
History's largest open model is free to download. Running it needs over 1.5 TB of VRAM. The real cost on AWS, GCP and Azure.
Auto-memory writes since v2.1.59, auto-dream cleans since v2.1.87. Demo: 14 files with months of contradictions reduced to 7 clean ones.
4 AI specialists coordinate themselves, Designer, Developer, Architect and SEO, and deploy a complete website to production. Zero code written.
Implementation guides for Agent Skills, Google ADK + A2A and Claude Agent Teams. Code, diagrams and demos.
Agentic Engineers: for people building real systems with agents.
I am Nicolás Neira García, a Software Engineer focused on Platform Engineering, automation and cloud: internal APIs, pipelines and service integration on Google Cloud and AWS, so teams can operate with more autonomy.
These days I embed AI agents inside those flows, agents that talk to real APIs, cloud services and pipelines. I share it here and on my channel from the actual implementation, with the terminal open.
I also run Agentic Engineers, a community for engineers building real systems with AI agents.