Two things happened this week that signal where AI agents are actually heading. OpenAI's Codex expanded beyond code to handle general computer use - 42% faster than before. And Claude started powering creative tools like Blender and Adobe applications, moving from text generation into visual workflows.
The pattern is clear: agents aren't just getting better at their original tasks. They're crossing boundaries.
Codex for Knowledge Work
Codex was built to write code. Now it's running entire knowledge-work processes. According to Latent Space's analysis, the latest version handles computer use tasks 42% faster than previous iterations - meaning it can interact with applications, navigate interfaces, and complete multi-step workflows without human intervention.
That's not about speed for its own sake. Faster execution means agents can handle more complex tasks within practical time constraints. A process that took 10 minutes might now take six. That's the difference between "interesting demo" and "actually useful tool".
The expansion into general knowledge work matters because it shifts Codex from a developer tool to a business tool. Legal research, financial analysis, customer support - these aren't coding tasks, but they involve navigating software interfaces, pulling data from multiple sources, and synthesising information. Codex can now do that work.
Claude for Creative Work
Claude's integration with creative software like Blender and Adobe tools represents a different kind of boundary-crossing. Creative work isn't structured like code or knowledge tasks. It's iterative, subjective, and often requires visual judgement.
The fact that Claude can assist in 3D modelling or video editing suggests the model has developed spatial and visual reasoning capabilities beyond text generation. It's not just suggesting steps - it's understanding what "good" looks like in visual domains.
For designers and creators, this means AI moves from being a writing assistant to a collaborator in visual workflows. That's a meaningful shift. The bottleneck in creative work is often execution - getting the idea out of your head and into the tool. If Claude can handle part of that execution, it changes how quickly ideas become real.
The Open-Weights Race
While the proprietary models grab headlines, the open-weights space is moving fast. Qwen3.6 27B just became the new leader in open-weights performance, outperforming previous benchmarks and offering capabilities that were proprietary-only six months ago.
This matters for developers and businesses that need control over their AI stack. Open-weights models can run locally, on-premise, or in private clouds. No API costs. No data leaving your infrastructure. For privacy-sensitive industries - healthcare, finance, legal - that's not a nice-to-have. It's a requirement.
The performance gap between proprietary and open-weights models is narrowing. Not closed yet, but narrowing fast enough that businesses are starting to make different architectural decisions. Why build on an API when you can run a comparable model yourself?
Scaling Trends and What They Mean
The Latent Space analysis also tracks scaling trends across model releases. The pattern isn't just "bigger is better". It's "smaller models are getting surprisingly capable".
We're seeing efficiency gains that weren't expected this quickly. Models with fewer parameters are achieving performance that previously required much larger architectures. That means lower compute costs, faster inference, and the ability to run powerful models on devices rather than in data centres.
For businesses, that changes the deployment equation. A process that required cloud GPUs last year might run on a local server this year. That's not theoretical - it's happening in production environments right now.
The Immediate Implications
What does this mean for people building with AI today?
First, agent capabilities are expanding faster than most roadmaps anticipated. If your 2025 plan assumes agents are limited to narrow tasks, it's already outdated. These systems are crossing into general-purpose territory.
Second, open-weights models are catching up fast enough that proprietary APIs aren't the only viable option anymore. That changes procurement decisions, security considerations, and long-term strategy.
Third, creative and knowledge-work domains are no longer separate from AI automation. The tools people use every day - Adobe, Blender, office software - are becoming AI-native. That's not a future trend. It's shipping now.
The agent era isn't coming. It's here. And it just expanded into territory most people assumed was years away.