Anthropic shipped Dynamic Workflows in Claude Code on May 28 alongside Opus 4.8, and the InfoQ writeup landed in early June. The capability: Claude dynamically writes orchestration scripts that fan out tens to hundreds of parallel subagents in a single session, with critic subagents that try to refute the findings, and the run keeps iterating until the answers converge. 16 concurrent agents run in parallel; cap of 1,000 agents total per run. Use cases Anthropic positions explicitly: investigating widespread bugs, managing large migrations, conducting security audits, performance reviews, and architecture analysis of complex software projects. Available in research preview today in Claude Code CLI, Desktop, and the VS Code extension for Max, Team, and Enterprise plans (if admin-enabled), plus on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.
The capability is a real shape-change versus prior Task-tool delegation. Before, you got one subagent per Task call; coordination across multiple subagents was your job to write, in your own loop, with your own message passing. Dynamic Workflows hands the orchestration script writing to Claude itself, and Claude composes that script with the concurrency primitive (parallel fan-out, up to 16 in flight at once) plus the convergence primitive (critic subagents that refute, iteration until converge). The "1000 cap per run" is a runaway-loop backstop, not a target, but it tells you the design intent: this is for jobs where the right number of subagents is unknown ahead of time and may be in the hundreds. The Jarred Sumner example Anthropic cited: 99.8% of the existing test suite passing across roughly 750,000 lines of Rust, eleven days from first commit to merge. That is the kind of migration that single-agent loops have not been able to close.
Two ecosystem threads. First, this changes what "agent" means in practical terms. The single-agent ReAct loop has been the implicit unit of work for two years; Dynamic Workflows reframes the unit as a multi-agent ensemble that decomposes-then-verifies. For builders thinking about their own agent platforms, the question is no longer "how do I make a better single agent" but "what is the fan-out-and-converge harness that produces results a single pass cannot." MiniMax M3 and Qwen3.7-Plus both gestured at agent-team patterns; Dynamic Workflows is the implementation that ships with the model that wrote it. Second, the verification-by-refutation primitive is the design choice worth pausing on. Most multi-agent frameworks use parallel for speed (cover more ground) or for ensembling (vote). Anthropic is using parallel for adversarial check: agents try to refute each other's findings before reporting up. That is the move that distinguishes "more agents" from "better answers," and it is the part most ad-hoc orchestration setups skip.
Monday morning, if you're on Max, Team, or Enterprise Claude plans: Dynamic Workflows is research preview today, worth trying on a real comprehensive-coverage task (security audit, large migration, full-codebase review) rather than on a small task where single-agent already works. If you're building your own multi-agent platform: the parallel-with-refutation pattern is the design lesson, and the 16-concurrent / 1000-cap shape is a sensible default to copy. If you're on the API path (Bedrock, Vertex, Foundry): same availability, same shape. Caveats: this is still research preview, so behavior may change; the 1000-agent ceiling is real, so workflows that try to spawn more will hit it; and cost scales with subagent count, so the "is the answer worth N subagents" question is now an explicit dial, not a hidden cost.