Opus 4.8 + Dynamic Workflows: 1000 agents for 1 prompt

Anthropic just shipped Opus 4.8 on May 28, 2026. Officially one of their best models to date. But the real news isn't in the benchmarks. It's what ships alongside: Dynamic Workflows. You give ONE prompt, and Claude launches up to 1000 agents in parallel, coordinates them, fact-checks their outputs, and brings you a result. With no human in the loop.

In this article: the real Opus 4.8 benchmarks, the Fast Mode divided by 3, the full Dynamic Workflows architecture, the Bun case study explained in detail, and how to activate all of this on your machine this week.

Opus 4.8: Anthropic's new model

Opus 4.8 ships less than two months after Opus 4.7. The pace is accelerating. Concrete changes for devs:

Official benchmarks

The most striking jump is on USAMO: +27 points in a few weeks, that's rare. On agentic coding (SWE-bench Pro), we go from 64% to 69% — confirming the trajectory holds.

Honesty x4

Anthropic emphasizes one key point: Opus 4.8 is about 4 times less likely than 4.7 to let a code flaw slip through without flagging it. For devs running Claude Code in autonomous mode, this is the real differentiator. Fewer "it works" claims that break in production.

Fast Mode divided by 3

Standard pricing stays the same: $5/M input tokens, $25/M output. But Fast Mode drops to $10/M input, $50/M output, that's 3x cheaper than the Opus 4.7 Fast Mode. For 2.5x the Standard speed.

Concretely: the barrier that held back long-horizon background sessions just dropped. You can finally leave Claude Code running for hours without panicking about the bill.

Dynamic Workflows: the real news

This is where the pattern shift happens. Not in the model, in how you use it.

Before Dynamic Workflows

• You code the multi-agent coordinator YOURSELF (CrewAI, LangGraph, custom scripts)
• You define: who does what, in what order, how to merge results
• You handle conflicts, retries, state management
• High setup cost. Reserved for multi-agent pros.
• 99% of devs stay sequential because it's too painful to set up.

With Dynamic Workflows

• You give the objective in natural language. Done.
• Claude writes the orchestration script itself
• It spawns N agents (10, 100, up to 1000 max per run)
• 16 concurrent agents simultaneously, max
• Adversarial check: 2 Claude reviewers per output, who try to refute the result
• Iteration until convergence (agents must agree)
• Resumable state on multi-day jobs — picks up after crash
• Available from Claude Code CLI + Desktop + VS Code extension via /workflows

The 5-step architecture

1. Bootstrap: you send 1 prompt — the objective in natural language
2. Fan-out: Claude plans and spawns up to 1000 agents (16 concurrent max)
3. Adversarial check: 2 reviewers per output try to refute
4. Convergence: iteration until agents agree
5. Final merge: result delivered + resumable state for multi-day jobs

The Bun case: proof at scale

Jarred Sumner, CEO of Bun, used Dynamic Workflows to port the entire Bun codebase from Zig to Rust.

The numbers

• Task: port Bun from Zig to Rust (750,000 lines of code)
• Setup: 1 initial prompt from Sumner
• Architecture: 2 chained workflows
  • W1: mapping Rust lifetimes for every struct field
  • W2: writing each .rs file as a behavior-identical port
  • Final fix loop: drives build + tests until clean
• Scale: hundreds of agents in parallel, 2 reviewers per file
• Result: 11 days from first commit to merge, 99.8% tests passing

It's the first time an agent successfully completed a codebase-scale migration in full autonomy. No human in the loop. The port is still research preview on Bun's side too, but the code is shipped.

When to use Dynamic Workflows

Not for everything. Overhead is high (burns 10x tokens vs a classic session). Reserved for big jobs.

Use for

• Massive codebase migration (language, framework)
• Refactor across 100+ files
• Multi-file security audit
• Porting an entire runtime (Bun case)
• Multi-angle research with consensus

Do NOT use for

• Fixing 1 bug
• Writing 1 function
• Modifying 1 file
• Quick web research
• Simple question

How to activate Dynamic Workflows today

# 1. Update Claude Code
claude update  # version 2.1.154+

# 2. Verify your plan (Max / Team / Enterprise required)

# 3. See Dynamic Workflows runs
/workflows

Opus 4.8 is already the default xhigh effort in the new version. First run: start small (10 agents) to validate before letting 1000 loose.

The pattern that's coming

Dynamic Workflows = the "dynamic orchestration" pattern. Before: workflow = static code (CrewAI, LangGraph, custom scripts). After: workflow = dynamic, generated by the model itself.

This is what OpenClaw, Hermes, n8n will copy in 3 to 6 months. Devin and Cognition already use variants. You'll never code a multi-agent coordinator by hand again. Your job becomes: describe the problem, verify the outputs.

Conclusion

Opus 4.8 is a model jump. Dynamic Workflows is a pattern jump.

The signal that we're entering the era where you describe your problem, and the AI decides how many agents to launch, how to coordinate them, how to verify the result. You become the architect again. Not the orchestrator.

If you're on Claude Max: you already have access. Test it on a small refactor this weekend. If you're on Pro: it'll trickle down in 3 to 6 months. The one thing not to do: ignore it.

Full breakdown in the video: https://mkc.sh/opus48?utm_source=blog

Grab the 14-slide premium pack (interactive HTML + high-quality PNG): https://mkc.sh/the-agentic-dev?lead=opus48-dynamic-workflows&utm_source=blog