AI · · 8 min read
Agents That Ask Before They Act: What October’s Releases Tell Us About the Next Year of Software
On 21 October Anthropic opened a research preview of Claude Duet, an app in which Claude works as a proactive agent in the background: it checks connected apps, brings you what needs attention, keeps working when you step away, and asks before it acts on your behalf. Five days earlier OpenAI updated GPT-6 Pro so a single task can be split across several agents working at once. On 13 October Google made Gemini 4 Argon broadly available after first restricting it to vetted cyber defenders, and on 7 October Anthropic released Claude Haiku 5.5 at ten cents per million input tokens. Four releases, one direction: capable agents, running continuously, cheap enough to leave on — and surrounded by explicit checkpoints.
The approval gate is now a headline feature
Duet is the clearest statement yet of how the labs expect agents to work. It will draft a message, prepare a payment or line up a file to be deleted — and then stop and show you what it intends to do. Sending messages, making payments, signing documents and deleting files all require explicit approval. Some actions, such as trading securities or moving money between accounts, it refuses outright. Anthropic is open that it can still misjudge what is urgent.
A year ago the pitch for agents was autonomy. In October the pitch is restraint: the agent is valuable precisely because it brings you a decision instead of making it. That is the same conclusion we reached building workflow systems for clients, and it is reassuring to see it become the default at the frontier.
Parallel agents change what “a task” means
GPT-6 Pro splitting a task across several agents at once sounds like a speed feature, and it is — but it also changes the unit of work. A request like “reconcile this month’s claims against the policy and flag the exceptions” no longer runs as one long chain; it fans out, works in parallel, and reassembles. The ceiling on what you can ask in one go rises accordingly.
For the business building on top of this, the implication is that review has to scale too. If an agent returns fifty proposed actions instead of one, the interface that lets a person approve, edit or reject them in bulk — and see why each was proposed — becomes the bottleneck. We now treat that review surface as a first-class screen, not a confirmation dialog.
Frontier capability reaches defenders first
Gemini 4 Argon’s staged release — vetted cyber defenders first, the public two weeks later — follows OpenAI’s decision in September to route its flagship model to its cybersecurity programme before general access. The labs are saying plainly that these models are strong enough to matter in security, on both sides.
Small and mid-sized Malaysian businesses should read that as a prompt, not a curiosity. Attackers will have access to the same class of capability. The basics — multi-factor authentication, least-privilege access, prompt patching, a tested backup, and software that validates its inputs — are no longer hygiene. They are the price of running anything connected to the internet in 2027.
How we are building for this
Every workflow system we scope now assumes an agent will eventually sit in it. That means three things from the first design: actions are classified as autonomous, approval-required or forbidden; every action is logged with its reasoning and its author, human or machine; and the review screen is designed for volume, because parallel agents will produce it.
Haiku-class models at ten cents per million tokens make it affordable to let such an agent run all day on routine work. The expensive part, as it has been all year, is the judgement around it — and that is where a business should put its attention and its budget.
Key takeaways
- October’s agents (Claude Duet, parallel GPT-6 Pro) are built around explicit approval before consequential actions — restraint is now the selling point.
- Parallel agents raise the volume of proposed actions; the human review interface must be designed to scale.
- Frontier models reach security teams first; SMEs should treat security basics as mandatory and build workflows with autonomous / approve / forbid classes and a full audit log.