Anthropic shipped Opus 4.8 and dynamic workflows the same day, and I made a video running one on a real job so you can see how it behaves, session limit and all. The CTO’s launch post said he was most excited about Claude planning work, fanning out across hundreds of parallel sub-agents in one session, and verifying its own output before handing it back. That is what I tested on camera.
The limits, stated plainly
A dynamic workflow is a JavaScript script that Claude writes and a runtime executes in the background. It can spawn up to 1,000 sub-agents per run, with 16 running in parallel. Because it runs in the background, the main session stays responsive. In the video I show the difference directly: fire off plain sub-agents and the session blocks, so a “hi” just queues up, but start a workflow and I can keep working while it runs.
What happened on the real run
I asked for a deep-research job: best local LLMs and inference engine for a 128GB Mac in 2026, compare engines and tokens per second. On the 5x plan I hit my session limit live, mid-run, and I left that in rather than re-record. The first attempt failed after burning about a million Opus tokens. Tokens are payroll on these, and a large workflow eats them fast, which is exactly the thing worth seeing before you fire one off.
On resume, Claude read its own script, fixed the root cause by changing how the search-fetch-verify agents did web work and dropping a schema constraint, then finished. The final run used 104 agents, not the full 1,000, pulling 22 sources and 85 claims, of which 25 were verified and 14 killed. The answer it returned: MLX is the fastest engine, Ollama the easiest, llama.cpp in between, and Qwen 3 Coder 30B the standout pick for coding.
When to use what
The most useful part of the video is the built-in decision guide, so here it is in text.
- Reach for a sub-agent when the work is bounded and fits your session, accepting that it blocks and Claude decides turn by turn what to run.
- Reach for a workflow when the job is large and parallel and should run in the background with results held in script variables, which is where the 1,000-agent ceiling matters.
- Reach for a skill when you want a reusable capability you invoke, a different concern from fan-out.
Set the effort to ultracode and Claude plans a workflow for every substantive task on its own. One cost tip from the run: workflows use your session model, so switch to Haiku or Sonnet instead of Opus when you do not need the top tier.
Watch the video for the live session limit, the token cost, and the self-repair. Anthropic’s docs: https://code.claude.com/docs/en/workflows
Subscribe for more.
I specialize in wiring multi-agent workflows into real engineering and research pipelines. Book a call at cloudyeti.io/meet.