Published: 2026-05-31

Claude Opus 4.8 Dynamic Workflows: How to Launch a 1,000-Agent Swarm

Chapters / key moments (click to jump — plays here on the page)

Claude Opus 4.8 shipped with dynamic workflows that let a single task spawn up to 1,000 parallel sub-agents, with reviewer agents that verify the work before results come back to you. Julian Goldie explains how to trigger ultracode mode, the five-step Goldie swarm stack for running big jobs end-to-end, and why Agent OS is the essential companion for coding to your own standards.

Source video

"New Claude Agent Swarms Update is INSANE!" by Julian Goldie SEO — Watch on YouTube →

Key Takeaways

Claude Opus 4.8's agentic coding benchmark score rose from 64.3% to 69.2%. Dynamic workflows support up to 16 concurrent agents and up to 1,000 total agents per workflow run.
Enable ultracode with /effort ultracode in your Claude Code session, or just type the word "workflow" anywhere in your prompt — Claude decides when a task is big enough to warrant a full swarm.
Dynamic workflows are on by default for Max and Team plan users; enterprise admins must enable them in Claude Code settings.
The Goldie swarm stack is five steps: Command (one English instruction), Swarm (parallel sub-agents), Verify (reviewer agents), Watch (live dashboard), Keep (organized output workspace).
Pair with Agent OS's /discover-standards before running a swarm — this feeds Claude your coding conventions so the entire swarm builds to your project's standards, not its best guess.

How Dynamic Workflows Work

You give Claude one instruction — "audit every API endpoint under src/routes/ for missing auth checks." Claude breaks the task into independent pieces, spawns parallel sub-agents to work each piece simultaneously, and assigns separate reviewer agents to try to find holes in the results. The run iterates until answers converge, then everything lands in a single organized output. Jarod Sumner used this to port the Bun runtime from Zig to Rust — hundreds of agents in parallel, two reviewers per file.

The system always asks for confirmation before kicking off because costs can be significant. Think of it as the right tool for large-scale code analysis, migration, or security sweeps — not for smaller day-to-day tasks where a regular session or sub-agent is faster and cheaper.

Installation and First Run

npm install -g @anthropic-ai/claude-code

Once installed, open your project and run claude. For your first workflow run, use a read-only task like a codebase audit — you get to learn the system without risk of unintended edits. Before running any swarm, let Agent OS discover your standards with /discover-standards and shape your spec with /shape-spec in plan mode.