AI is moving quickly enough that a list of headlines is no longer very useful. The interesting question is what is actually changing underneath the headlines?
This week, three themes stand out: AI agents are gaining more ability to act beyond a chat window; video models are collapsing more parts of the production workflow into one model; and the infrastructure around agents is becoming much more practical.
AI agents are moving beyond the browser
The biggest shift is not simply that models are getting better at answering questions. They are increasingly being connected to tools, systems and real-world interfaces.
OpenAI's GPT-6 Astra, released to a limited set of organisations this week, is explicitly designed for complex multi-step work, including coding, research and computer use. OpenAI says Astra has reached its “Critical” cybersecurity capability threshold and has added stronger monitoring and safeguards for agentic use.
Elsewhere, the agent ecosystem is expanding beyond traditional computer-use workflows. Doop lets AI agents work alongside people on a shared design canvas through MCP, while Monid is building a tool layer that lets agents discover and call more than 1,700 tools and APIs. In India, the next step could be even more tangible: Reuters reports that NPCI is preparing a framework that could let AI agents make low-value UPI payments under defined rules, limits and identity controls.
AI video is getting closer to an all-in-one workflow
AI video is also moving away from the idea that generation, audio and editing have to be separate steps.
MiniMax's H3 is a good example. The model accepts text, images, video and audio as multimodal context and can generate video with native stereo sound at up to 2K resolution and 15 seconds per clip. The important part is not just resolution: the model is designed around the relationship between the visual and audio content, rather than treating sound as an afterthought.
That direction matters because a usable video workflow is more than a pretty clip. Creators need dialogue, ambience, sound effects, references, consistency and the ability to iterate without rebuilding every part separately.
The agent infrastructure layer is filling in
Agents need more than a capable model. They need context, permissions, tools, APIs, deployment environments and a reliable way to execute tasks.
That infrastructure is starting to become a product category of its own. Zoho Catalyst 3.0, announced this week, combines agent skills, MCP access and a full-stack cloud environment so an agent can work with databases, authentication, functions and deployment tooling. Zoho reports that its internal tests saw task completion rise to roughly 90–95% when models were given the platform's skills and MCP access, with human intervention falling close to zero on the tested workflow.
The broader pattern is important: instead of asking an AI model to generate instructions for a human to follow, platforms are exposing structured operations that an agent can actually call. That is the plumbing required for agentic software to move from demos into repeatable workflows.
The bigger picture
Three things are happening at once:
- Models are becoming more capable of acting.
- Interfaces are becoming more multimodal and connected.
- Infrastructure is becoming more agent-ready.
That combination is what makes the current AI cycle different from another round of chatbot upgrades. The hard part now is not only generating intelligence — it is deciding where that intelligence can act, what it is allowed to touch, and how we keep humans in control.
🫀 Previous Pulse Checks
New editions will be added here each week. The archive will build as the series grows.
OpenAI — GPT-6 Astra safety overview · MiniMax — H3 · Zoho — Catalyst 3.0 · Doop — MCP / agent canvas · Monid — agent tools & APIs · Reuters — agentic payments on UPI