summree
Claude just dropped UltraCode... its Insane
Claude
Jack Roberts

Claude just dropped UltraCode... its Insane

⏱ 20 min video · 3 min read9 Jun 2026
TL;DR
Jack Roberts demos Claude Code's new UltraCode feature, which spins up to 10 parallel AI agents managed by deterministic code rather than a master agent, delivering faster and more accurate outputs. He shows how to use it correctly, when NOT to use it (token costs are 4-7x higher), and builds a live model-comparison dashboard into his agentic OS as a practical example.
Key points
1
UltraCode is activated by typing 'ultra code' in the prompt — it is NOT the reasoning effort slider (low/medium/high/max) in Claude Code
2
Instead of an LLM managing sub-agents (which can forget, hallucinate, or become confidently incorrect), UltraCode uses deterministic JavaScript/code to orchestrate up to 10 parallel agents, guaranteeing it runs until the desired output is achieved
3
Token costs are 4-7x higher with UltraCode due to parallel fan-out, so it should only be used for roughly 20% of tasks — research, audits, multi-perspective reviews, and separable systems
4
UltraCode is NOT suitable for single-file refactors, bug fixes, quick sub-5-minute tasks, or any strictly sequential step-by-step workflows where Agent A must pass to Agent B then C
5
A live demo builds a model-comparison dashboard (pulling data from OpenRouter) into an agentic OS, letting users filter AI models by speed, cost, benchmarks, and popularity to inform routing decisions
Actionable insights
Always have a clarifying conversation with Claude before triggering UltraCode — the system is only as good as the task definition, so poorly scoped prompts waste expensive parallel compute
Use 'ultra code' for a single turn (surgical) or 'ultra code on' for an entire session, but reset when you exit — never leave it on by default for casual conversation or trivial edits
Apply UltraCode specifically to: adversarial claim verification, judge-panel debates, multi-perspective research, and building reusable separable systems where you do not yet know the shape of the answer
Notable quotes

It's like firing a bazooka when a hammer very easily would have done the job for us.

Lock-in is the enemy of high performance — the winners are going to route by task.

My wallet is crying watching the video — it is very powerful, but you don't want to be using it all the time for things where it is not required.

Worth watching?
⏭️
Worth watching the full video?
The core concepts, decision framework, and use-case boundaries are all captured here — only watch the full video if you want to see the live UltraCode demo and dashboard build-out in action.
Topics
AI & TechClaude

Explore more summaries on these topics →

Saved you some time? The creator still deserves a like.

Watch on YouTube →
More like this

Want this for your own channels?

Add the channels you follow. Every new video summarised and in your inbox the moment it drops. From £4/month.

Try it free