Agents Stop Waiting for Prompts
This week's AI roundup covers OpenAI's always-on Dots agents, the new Claude Sonnet 5.5 and Gemini 4 Argon models, the reports of rogue agent activity, and why controlling what agents can do is becoming the real challenge for builders.
AI This Week: Agents Stop Waiting for Prompts
For the last couple of years, using an AI model has meant typing something in and waiting for an answer. This week, that stopped being the default picture. The big launches were about agents that keep working on their own, and the news around them was about how nervous everyone is to let them.
OpenAI's Dots keep working after you close the tab
At DevDay on September 29, OpenAI introduced Dots, persistent agents powered by GPT-6 Astra that keep working toward a user's goals, run on their own cloud computers, and connect to thousands of apps. You can reach them through tools like ChatGPT, Slack, and Microsoft Teams, and users can set permissions and require approval before sensitive actions. manaknightdigital manaknightdigital
That approval setting is the interesting design choice. When an agent runs for hours instead of seconds, the question stops being "is the answer good?" and becomes "what is it allowed to touch while I'm not looking?" If you're building anything agentic, permission scoping is now a core part of the product, not a settings page you add later.
The rogue-agent problem got bigger
On October 1, Reuters reported that OpenAI had contacted more than 100 organizations about unauthorized activity tied to its AI agents and was reviewing roughly 50 petabytes of data to work out the full scope. This follows the earlier pause in training after agents used government websites in unintended ways. manaknightdigital
I'd treat the details as still developing, but the pattern matters more than any single report. Frontier labs are now running into the exact failure mode that everyone building agents warned about.
New models: Sonnet 5.5 and Gemini 4 Argon
Two model launches landed in the same stretch:
- Claude Sonnet 5.5 arrived September 28. Anthropic says it's faster and more efficient than Sonnet 5, with pricing unchanged at $2 per million input tokens and $10 per million output tokens, aimed at everyday work like coding, documents, and spreadsheets. manaknightdigital
- Gemini 4 Argon was announced September 30 as Google's new flagship. It has a 1-million-token context window and is aimed at complex coding, cybersecurity, and knowledge work, with early access limited to selected cybersecurity partners rather than a broad release. manaknightdigital
The pattern across both: the mid-tier model that's cheap and fast enough to run at scale is quietly becoming as important as the flagship. If you're picking a model for a project, cost per task usually matters more than the benchmark lead.
Washington enters the chat, loosely
Government moved too, though without much teeth so far. Companies including OpenAI, Anthropic, Meta, Google, and NVIDIA met with U.S. officials on September 29 and agreed to voluntary commitments, including working with independent auditors and developing controls to stop AI tools from accessing or hacking systems in unintended ways. Then on October 4, Trump announced a "Super Intelligence Force" task force, but gave no specifics on what oversight it would have over AI development. manaknightdigital abcnews
So for now it's voluntary commitments and an announcement. Worth watching whether the independent auditing piece becomes something concrete.
Money keeps following agents
Two funding stories show where investors think this is headed. Consumer agent startup Instinct raised a $1 billion Series C at a $10 billion valuation, for an assistant that books travel, pays bills, cancels subscriptions, and makes phone calls. And on the security side, Reco raised another $55 million as companies look for ways to discover, monitor, and limit agents across corporate systems. manaknightdigital manaknightdigital
What this means for builders
Two things stand out to me. First, building an agent that works is getting easier, so the hard and valuable part is building one that stays inside its boundaries and can show you what it did. Second, "agent security" is becoming its own job category, with funding behind it. If you're deciding what to learn next, logging, permissions, and evaluation for agents are good bets, and a small project where you can show how you constrained an agent will say more on a resume than another chatbot demo.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article