Signals · Capability Shift
Google's Gemini 4 Carbon splits work into agent roles, and the judgement roles are the ones to hold
Google's Gemini 4 Carbon hands agents research, triage and drafting, and the roles Google suggests for its strongest model show where judgement now sits.
by Jo·3 min read·
New here? Start with the free AI Survival Kit →
Google's Gemini 4 Carbon splits work into agent roles, and the judgement roles are the ones to hold
0:00 / 5:32
This voice is generated by AI.
Part of the guide Will AI take your job? How to find out for your role, in an afternoon
Google released Gemini 4 Carbon to developers, enterprises and paying subscribers on Tuesday, in a post on its company blog by Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect. Google calls Carbon its strongest model, built for "long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning". It says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at "more than 2X lower cost per task" than the next best external model. Every figure here is Google's own, and several come from internal benchmarks. Read the launch first for the work it describes, then for the work it leaves with people.
The org chart is in the product
The most useful line in the post is about a feature. Google is previewing /swarm in its Antigravity tool, which turns "a single prompt into a coordinated team of agents". Here is the setup Google suggests. Give Carbon the roles that "demand the deepest reasoning, like architectural design, coordination, and auditing". Pair it with faster, lighter models working in parallel. Add "adversarial verifiers that test every change".
That is a team structure. A lead, a group of executors, a review function. Google has written the division of labour for a software team into a feature menu.
The case studies show how far down the task list this reaches. Google says that in Wiz's Scan for Good programme, Carbon found 0-day vulnerabilities at two major SaaS providers in under two hours, "compressing weeks of vulnerability research to a single afternoon". Google researchers used it with a formal verification framework, AlphaProof Code, to prove key security properties of the code behind Signal's cryptographic protocol in weeks. Google says that task "historically takes expert teams years". For legal, finance and enterprise operations, Google is selling synthesis across large sets of dense documents, plus polished spreadsheets, presentations and reports. The introductory price is $2 per million input tokens and $10 per million output tokens.
Those are tasks. They are also the tasks that fill the first years of many careers: first-pass research, the draft deck, the triage queue.
What stays with the person
Now read the same case studies for who did what. Google's Cloud Vulnerability Research team used Carbon to build a proof of concept, then helped the maintainers reclassify the bug as critical and issue a patch. Wiz worked with the affected organisations to fix the flaws. Google's researchers, using Carbon alongside other models, turned the security goals into mathematical theorems before AlphaProof Code agents wrote the Lean proofs.
In the security cases, people at Google and Wiz took the model's findings to the affected maintainers and organisations. AI does the routine work. You do the thinking. Carbon moves the line of what counts as routine a long way up the task list. Some of Google's examples name the people who acted on the result; others, such as Project Fortify, are described as autonomous.
The roles Google suggests giving the strongest model are architecture, coordination and auditing. Those are the roles worth holding yourself, with the models working for you.
Your Next Move
- Run the comparison on your own week. List last week's tasks and mark the ones that match Google's pitch: document synthesis, first-draft reports, triage, research loops. Our step-by-step method for working out whether AI will take your job covers the rest. Whatever stays unmarked is where to build.
- Practise the auditor role this month. Take one piece of work you would normally produce yourself and have a model draft it. Write down every error you catch and why it matters. That record is evidence of judgement, which is what an employer pays for once drafting is cheap.
- Treat the benchmarks as claims. Google's numbers are Google's, and several come from its own tests. Before your team adopts Carbon or a rival, run it on three real tasks from your own queue. Compare the cost per finished task, with you checking the output.
I build systems like the one publishing this site. → Work with me
About the author
Jo
Jo runs The War Room: one signal a day on how AI is changing work, and what to do about it.
Sources
Primary source
Get the Briefing
Want more intelligence like this?
The weekly briefing, free — and the AI Survival Kit with it.
More in Signals
See allSignals · Capability
Siemens CEO Roland Busch says AI could transform society in 7 to 8 years, and AI skills will decide who gains
3 min read
Signals · Power
Windows 11 is recording which AI coding agents you run, and in one US test the upload setting was on
3 min read
Signals · Labour Market
Pay moves to the bottleneck. A careers list that is really a map of who owns the work
3 min read