Articles/Agents
Subagents Went Mainstream in 2026. The Hard Part Is the State Between Them.
Claude Code, Claude Managed Agents, Codex, Gemini and Letta all shipped multi-agent orchestration this year. What the pattern looks like and where coordination breaks.
Jose R. Gonzalez, CPA · Founder of ContxOS
Updated · 2 min read
In short
The lead-agent-plus-specialists pattern went from advice to product in 2026. Claude Code added experimental Agent Teams in February, Claude Managed Agents put multiagent orchestration in public beta in May, Gemini’s managed agents expanded in July, and Letta shipped agent-written orchestration in September. Every version faces the same problem: what state the lead passes to a subagent, what comes back, and where the overall position is recorded, so the work survives when any one agent’s context ends.
From recommendation to feature
- February 2026: Claude Code shipped Agent Teams as a research preview, with a shared task list and messages between teammates, and gave subagents a persistent memory field.
- May 2026: Claude Managed Agents put multiagent orchestration into public beta. A lead agent delegates to specialists with their own models, prompts and tools, on a shared filesystem.
- July 2026: Google expanded Managed Agents in the Gemini API, with remote MCP support.
- September 2026: Letta’s dynamic workflows let agents write their own subagent orchestration scripts.
Where coordination breaks
- Too much context down: a subagent handed the lead’s whole history loses the benefit of a clean window.
- Too little back: a condensed result that leaves out what was verified forces the lead to redo the check.
- No durable position: if the lead’s window ends, the record of which subtasks finished is lost with it.
What it means in practice
Give each subagent only its step and the facts that step needs, have it report what it verified, and record the run’s position somewhere outside every agent’s context. ContxOS workflow runs work this way. A step that has proven reliable is dispatched to a subagent with just that step and its facts. The subagent attests its result to the run. The lead reads the run’s position before deciding what comes next.
Sources
- Claude Code changelog — Anthropic on GitHub
- Building a C compiler with a team of parallel Claudes — Anthropic Engineering, February 5, 2026
- New in Claude Managed Agents — Anthropic, May 19, 2026
- Expanding Managed Agents in the Gemini API — Google, July 7, 2026
- Dynamic workflows — Letta, September 28, 2026
Frequently asked questions
When is a subagent worth it?
When a step is self-contained, would fill the lead’s context with detail, and has a result that can be checked. Keep judgment calls with the lead.
Give your agent a memory that lasts
Connect ContxOS to Claude in about a minute. The free plan includes one project.
Product names belong to their owners. ContxOS is not affiliated with the companies named here.