The multi-agent pattern inside Claude Code did not arrive with the 2026 agent teams launch. It started as a small, boring feature in July 2025, and the arc from there to today is worth reading as one line.
July 24, 2025. Anthropic added custom subagents to Claude Code. The news was quiet at first. A Japanese developer site, gihyo.jp, reported the Custom Sub Agents feature on July 24, 2025, and egghead’s AI Dev Essentials newsletter logged the same release the same day. Wider coverage followed on July 25 and 26, when WinBuzzer described subagents as specialized AI assistants with independent contexts for each task.
What a subagent actually was: a project or user scoped specialist with its own context window, its own system prompt, and its own tool permissions. The main session delegates a task, the subagent runs it in isolation, and only the final summary returns to the parent thread. The canonical example, from the public issue tracker, was a code reviewer subagent that only reads files and runs tests.
The economics were the real story. A subagent keeps research noise out of the main context window, parallelizes independent tasks, and gives you a fresh reviewer that has not inherited your assumptions.
February 5, 2026. Opus 4.6 launched with agent teams, and the subagent primitive grew into orchestration. The stress test came from Anthropic researcher Nicholas Carlini before release: sixteen Opus 4.6 agents, no internet access, Docker isolation, tasked with writing a Rust based C compiler from scratch, able to compile the Linux kernel. Roughly 2,000 Claude Code sessions over two weeks, about $20,000 in API costs, around 2 billion input tokens and 140 million output tokens. The result is a 100,000 line dependency-free compiler on GitHub with x86, ARM, and RISC-V backends, a 100% pass rate on the GCC torture test suite, and real projects it compiles, including Linux 6.9, PostgreSQL with all 237 regression tests, SQLite, and DOOM. The Register published a skeptical follow-up two days later. Both takes can be true: the compiler is a stress test artifact, not a product, and the multi-agent loop is what actually held together.
May 28, 2026. Opus 4.8 added dynamic workflows, where a lead agent distributes work instead of the human writing every delegation. Reporting from the-decoder on Anthropic’s Managed Agents describes a lead agent distributing tasks across up to 1,000 sub-agents at once, with a test where a single agent found at most 27 of 70 hidden bugs in a codebase while the multi-agent workflow consistently found 66. That comparison is single-sourced, so treat the exact numbers as reported, not verified.
Sixteen months, three stages, one direction. The unit of work moved from one prompt answered by one agent to a coordinated team with a manager, specialists, and a budget. What has not settled is the part builders care about most: verification. More agents caught more bugs in the Managed Agents test, but the cost per finding is still the number nobody publishes first.