Chapter 9. Production Case Studies 2025–2026
By mid-2026, are no longer demos: Anthropic's multi-agent system outperforms a single agent by 90.2%, Claude writes more than 80% of Anthropic's own production code, and an agent takes part in more than one in five code reviews on GitHub. Yet the picture has two hard constraints: economics (multi-agent burns ~15× the of a chat) and a high failure rate — 74% of enterprises rolled back live customer-facing agents. So the key skill when reading this period's case studies is telling independently verified figures from vendor marketing.
Anthropic Research: The Reference Orchestrator-Worker
The canonical published case is Anthropic's multi-agent research system (June 2025). An architecture: a lead on Claude Opus 4 decomposes the query and spawns 3–5 on Claude Sonnet 4 in parallel, each with its own ; a separate CitationAgent places citations at the end. The result: +90.2% over single-agent Opus 4 on an internal research eval, while parallel plus parallel subagents cut research time by up to 90% on complex queries.
The price and the mechanism of success are one and the same: an agent spends ~4× the of a chat, a ~15×, and token consumption alone explains 80% of the quality variance (with the number of tool calls and the choice of model — 95%). Hence Anthropic's explicit economic rule: "Multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance." An important honest caveat from the same post: multi-agent is good for parallelizable "read" tasks and works worse on tightly coupled tasks like coding. Cognition, in Don't Build Multi-Agents (the day before Anthropic's post), showed this by example: parallel subagents without shared context make conflicting implicit decisions. The field's synthesis (Harrison Chase, LangChain): "read" swarms are simpler than "write" swarms — reads parallelize, writes serialize.
The "Claude writes Anthropic's code" case stands apart — two distinct, often conflated here. Boris Cherny: more than 90% of the Claude Code team's code is written with Claude Code (via WorkOS); and the company-wide claim: >80% of the code merged into Anthropic's production by May 2026 is authored by Claude (an Anthropic Institute paper, via The Next Web). Caveats from the same source: every change passes an automated Claude reviewer, and "human code review has become the constraint." The viral "100% of Anthropic's PRs are driven by Claude Code" is an unverified retelling from a tweet.
This is where the open part of the chapter ends. The rest opens with a free registration — no payment, just an email address.