Video: "Claude Opus 4.8 Ultracode is INSANE!" by Julian Goldie on YouTube.
What Ultracode actually is
Previous versions of Claude Code gave you one model working through a task sequentially. Ultracode changes the shape of what's possible: when you flip to high effort, Opus 4.8 acts as an orchestrator — it writes its own orchestration scripts and uses them to spin up a swarm of parallel sub-agents, each handling a different piece of the work simultaneously.
The result is that complex tasks that previously required you to prompt in stages — write the plan, check it, prompt again for execution, review — can now run from a single goal. Claude plans the work, distributes it, collects outputs, and compiles results. You still need to review and iterate, but the model does more of the mechanical sequencing on your behalf.
In practice, this is the same orchestration pattern that people have been building manually using Hermes Agent's swarm mode. Ultracode bakes a version of it into the Claude Code environment directly.
The other changes in Opus 4.8
Ultracode gets the headlines, but Opus 4.8 also arrives with effort controls across the board — low, medium, and high settings — so you can dial back the model's depth on simpler tasks and keep costs in check. There's also a cheaper fast mode for lighter inference work.
Anthropic made calibration improvements too. Opus 4.8 is measurably less likely to confidently give you a wrong answer, and more willing to say it doesn't know. That's a quieter change than Ultracode, but in practice it matters more for business use — hallucinated facts in a published document cause real problems.
Worth knowing: this is 41 days after Opus 4.7. Anthropic are releasing at pace. If you're building systems around a specific model version, plan for the cadence.
What it means for someone already using Claude Code
If you've been using Claude Code for coding or content tasks, Ultracode is an option you can try on any task you'd normally break into multiple prompts. Give it a goal rather than a step, and let the orchestration mode figure out the sequencing. It won't always be faster — orchestration has overhead — but for genuinely complex multi-step jobs, the time saving is real.
The honest caveat is that more parallelism also means more opportunities for errors to compound. A sequential workflow lets you catch problems at each step. A parallel swarm can get four things wrong simultaneously before you see any output. Review the results with the same care you'd apply to any delegated work.
What's still being figured out
Dynamic workflows are powerful when the goal is well-defined. The model is good at breaking down coding tasks, research tasks, and content pipelines into parallel streams. It's less effective when the goal itself is ambiguous — Ultracode will still produce output, but it'll be confident output heading in the wrong direction. Spend more time on the initial prompt, not less.
Julian's take is that Opus 4.8 is a genuine step up from 4.7 for agent work. That said, if your current Claude Code setup is working well, this is an upgrade rather than a rebuild — you don't need to overhaul anything to benefit from it.
Where this connects to NordSys
We use Claude Code as a core part of the agent stacks we build for clients. Opus 4.8's Ultracode mode changes what's practical on complex multi-step workflows — the kind of automated pipelines that previously needed daily prompting to move forward. If you're curious how that applies to your business, it's worth a conversation.
See our AI Agents service →