Architecture

Anthropic published one of its more technically substantive engineering blog posts this week: a deep dive into Claude Managed Agents, their hosted service for running long-horizon AI agents. The core thesis is elegant and directly relevant to anyone building production agent systems today. The Brain/Hands Problem The central challenge Anthropic addresses is one that every serious agentic AI practitioner has run into: your harness — the loop of code that calls Claude, handles tool results, manages context, and decides when to stop — encodes assumptions about what the model can and can’t do. The problem? Those assumptions go stale as models improve. ...

At RSAC 2026, four separate keynotes from four separate companies arrived at the same conclusion without coordinating: zero trust must extend to AI agents. Microsoft, Cisco, CrowdStrike, and Splunk each named AI governance as the biggest gap in enterprise security. The problem, as Cisco’s Matt Caulfield put it, isn’t just authenticating an agent once and letting it run — it’s that “at any moment, that agent can go rogue.” Now two vendors have shipped architectures that actually try to solve it. Anthropic and Nvidia have each published zero-trust AI agent frameworks — and they solve the credential isolation problem in fundamentally opposing ways. ...

One of the hardest problems in agentic AI development isn’t building a single capable agent — it’s designing systems where multiple agents coordinate effectively without breaking down. Anthropic has stepped into that gap with a comprehensive published framework covering five multi-agent coordination patterns, giving developers an authoritative reference for the architectural decisions that matter most when building complex autonomous systems. Why Coordination Patterns Matter A single agent can handle a bounded task well. But real-world complexity quickly exceeds what any single agent can do in a single pass — tasks require verification, parallelization, specialization, and feedback. The naive approach (just make one very smart agent) breaks down at scale: context windows fill up, error rates compound, and there’s no mechanism for catching mistakes before they propagate. ...

Google’s Agent-to-Agent (A2A) protocol is fast becoming the standard handshake for cross-agent communication in production agentic systems. A new deep-dive from freeCodeCamp — paired with a working GitHub implementation at win4r/openclaw-a2a-gateway — shows exactly how to wire A2A into an OpenClaw plugin so your agent can receive tasks from any A2A-compliant caller. This how-to summarizes the architecture and key implementation steps. For the full guide, see the freeCodeCamp article. What You’re Building An A2A plugin bridge does three things: ...

Somewhere in your company’s recent strategy deck, there’s a slide about multi-agent AI systems. It probably has a diagram with six or eight boxes connected by arrows, each box representing a specialized agent — one for research, one for synthesis, one for outreach, one for quality control. It looks clean. It looks powerful. It looks exactly like the microservices architecture slides that were circulating in 2014. InfoWorld is issuing the same warning now that engineers were quietly issuing then: distributed complexity is not a free upgrade. You have to earn it. ...

If you’re building production AI agents in 2026, you’ve almost certainly encountered both MCP (Model Context Protocol) and Agent Skills as architectural options. Both are ways to extend what an AI agent can do — but they operate at fundamentally different levels of abstraction, and choosing between them (or combining them) is one of the most consequential architectural decisions you’ll make early in a project. This guide breaks down how each approach works, when each excels, the compatibility patterns for using both together, and the production deployment tradeoffs that practitioners are discovering in the field. ...

The memory problem in agentic AI is well understood: most agents are stateless. They start fresh every session, have no record of past decisions, and can’t explain why they did something three interactions ago. For demos, that’s fine. For production systems that need to audit, adapt, and coordinate over time, it’s a serious architectural gap. Context graphs are one of the most architecturally interesting answers to that problem — and FalkorDB’s recent technical breakdown is worth understanding even if you don’t use their specific product. ...