Hisham AlhussainOctober felt like every AI company decided at once that it was time to own the agent platform. OpenAI's DevDay brought AgentKit, with a visual Agent Builder, plus the Apps SDK for building apps inside ChatGPT. A week or so later Anthropic launched Agent Skills, along with Claude Haiku 4.5. And that's before counting all the smaller launches in between.
I spent a good part of the month trying things out. Here's where I landed.
Visual builders: great for some things
I went in skeptical about drag-and-drop agent builders, and I came out... a little less skeptical. They're genuinely good for:
- Prototyping fast. Getting a flow in front of a stakeholder in an afternoon is valuable.
- Simple, linear workflows. Classify, look something up, respond. A canvas works fine for that.
- Non-developers. Letting an ops team tweak a flow themselves without filing a ticket is a real win.
Where they fall down for me is anything with real complexity. Branching logic, error handling, versioning, code review, testing. All the stuff that keeps production systems healthy is just easier in code. I've been burned before by "low-code" systems that turned into giant, untestable flowcharts nobody wanted to touch.
Skills are the one that stuck
Agent Skills surprised me the most. The idea is simple: a folder with instructions, and optionally scripts and resources, that the model only loads when it's relevant. It's basically packaged know-how.
Why I like it:
- It's good context engineering. The model sees a short description of each skill and only pulls in the full details when it needs them. That keeps the context window lean.
- It's just files. I can version them in git, review them in a PR, and share them across a team.
- It captures the stuff that's hard to prompt. Our house style for reports, the exact steps for a deployment, how we format a compliance summary. Things that used to live in someone's head.
Haiku 4.5 is quietly a big deal
Smaller, cheaper, fast models getting this good matters more than people think. A lot of agent steps don't need the biggest model. I've started using small models for the routine steps and saving the heavy hitters for planning and the hard calls.
Where I've landed
My setup right now: code for the agent loop and anything that needs testing, Skills for packaging domain knowledge, MCP for connecting tools, and visual builders for quick prototypes that may or may not graduate to code.
The platform race is going to keep heating up. My advice is to keep your core logic portable so you're not locked in when the next big launch comes along. Because there will be a next one, probably next month.