Your Multi-Agent Architecture Is Your Org Chart

TL;DR

ANTHROPIC SAYS FOLD YOUR SUBAGENTS: their applied AI team found single-agent loops with progressive tool disclosure outperform multi-agent harnesses on quality, cost, and latency in enterprise deployments. One customer measured a roughly 2x faster runtime. Claude Code's own system prompt shrank 80% via the same approach. Migration kit available. Decision matrix included for when subagents actually earn their place.
WEACT: LORA WEIGHTS AS TOOL CALLS: a frozen backbone runs capability-specific adapters the agent loads and unloads mid-task like any other tool. +27.5% over the best static LoRA merge on 8 deep-research benchmarks. Modular continual learning without touching existing weights. Code and data released.
Tool of the Week: AWS Infinity: 75,000 concurrent agents in 8GB RAM on a Raspberry Pi 5, running on stackless coroutines. When a tool call fires, Infinity serializes the full conversation state and suspends the process until results arrive, holding no thread and pinning no memory while it waits. Apache 2.0, research preview, AWS and UC Berkeley.

The Big One

Your Multi-Agent Architecture Is Your Org Chart. That's the Problem.

Source: Anthropic / claude.com | Published: October 23, 2026

The org chart became the architecture. One team, one subagent. A platform team added an orchestrator on top of that, every team contributed a sentence to the system prompt, and what started as a reasonable decomposition became a Frankenprompt with conflicting rules and a customer experience that 14 agents could not keep consistent.

Read the full issue

Subscribe to The Agentic Engineer to unlock this and every future issue.