Computer-use agents get cheaper, open models gain, designers push back

Agents move from demos into specific jobs

Several launches move agents out of demos and into particular kinds of work. Google extended its computer-use API to Gemini 3.5 Flash, so the cheap, high-volume model can now see a screen and click and type. It does this at lower cost and latency than the larger models that previously had the capability. Perplexity's Computer for Counsel builds on its Computer Use foundation to take on legal research and document tasks. Cisco is adding AI troubleshooting for industrial networks, applying agentic diagnostics to factory-floor settings where downtime is especially costly.

The rules for how agents may use other businesses' services are less settled. A detailed look at Amazon v. Perplexity examines how Perplexity's agent interacts with Amazon's site. It also covers the legal theories in dispute, including unauthorized access, scraping and enforcing terms of service against autonomous agents. 0x has set explicit terms for its own service: agents can now call its swap API for $0.01 per request in USDC, paying through the x402 scheme instead of API keys and invoices. For teams building shopping or browsing agents, the lawsuit shows where one large platform is drawing enforcement lines. 0x shows a service choosing to invite agent traffic and bill it per call.

Open models get serious about agent workloads

Interconnects argues that Zhipu's GLM-5.2 is a step change for open agents. It cites better tool use and multi-step reliability than earlier open releases, though it says the model narrows the gap with frontier closed models rather than closing it. That matters to teams that want agent stacks they can self-host or fine-tune. Evaluation is also moving past static tests: the Qwen-AgentWorld paper introduces a framework that tests agents across varied simulated tasks, stressing planning, memory and recovery from failure. For customization, NVIDIA describes techniques for faster fine-tuning of Hugging Face models with NeMo AutoModel, aimed at shortening long, costly training cycles on its hardware.

Designers resist handing design systems to agents

Three essays question how far agents should reach into design work. One critic argues that agentic design systems misunderstand the job, because design systems exist to encode shared judgment and constraints across teams, not just to generate components. Penpot makes a related case that tools like Anthropic's Claude Design handle surface-level output while skipping governance, consistency and collaboration, the problems that define enterprise design. A WebDesignerDepot piece takes a more constructive angle. It describes an emerging Agentic Experience (AX) designer role focused on how agents behave, negotiate and hand control back to people as more products are used through agents.

Talent moves, a mystery listing and misuse detection

TechCrunch reports that more Gemini researchers have left Google for Anthropic, continuing a pattern of researchers moving between the two labs even as Google keeps shipping. Separately, a viral thread says a model called Fable 5 has reportedly reappeared in Amazon Bedrock, with no official announcement explaining its origin. On the safety side, a writeup on Anthropic's Claude Compliance API argues that misuse detection must examine the content of requests and outputs, not only account or pattern signals. That makes it a practical reference for teams building their own trust-and-safety layers on LLM APIs.