We're quickly moving into an era where creating agent workflows will become more accessible and easier to create and manage. There's also a growing demand for education on how to safely and accurately use these tools, from deep integration into development workflows to business automations and personalized assistants.
The latest release of Claude Sonnet 4.5 has been impressive for coding tasks, and with the right scope of project rules, guiding IDEs is becoming more predictable with human oversight. It's less about "one shot" prompts and more about predictable guardrails. Incorporating predictable or creative design patterns only further enhances this process - such as using prebuilt components from 21st.dev to give UIs a less “vibe coding” look.
The best result as I reflect on these advancements is the ability to finely tune prompts using agent input and documentation as a guide. I can see that in the near term these tools will evolve into ways where we create very specific agents that hand-off tasks to other agents in a more predictable process and workflow. Human in the loop will be paramount in this progression, particularly with a keen eye on security.
In relation to security awareness, there was a security vulnerability found within the Figma MCP server where an attacker on the same network or via DNS rebinding could trigger tools like get_figma_data with crafted inputs containing shell meta-characters, leading to code execution, data exfiltration, or lateral movement from a developer machine running the MCP.
There are so many new AI tools to begin to explore for helping to create these agent workflows and automations. As you go, pick the more predictable processes and begin to leverage these resources across your development and creative teams. For example, you could train an agent on your codebase and documentation - API metadata where it can become an extension of your learning process and onboarding for new talent.
It's an amazing time to begin to leverage these tools from the IDEs to the no-code platforms and start to deliver personalized and rich user experiences. Build and create responsibly, stay local and sandbox these tools as you go. Pick the development patterns that are easiest to manage and deploy, giving you more predictable results.
One thing is for sure: things are moving quickly and patterns and deployments of these tools are changing almost faster than people can learn them. I'd recommend learning from the open source tools that exist today and adopting them into your day-to-day processes to enhance your work and adapt only the tools that allow you to give them the oversight they need.