Hisham AlhussainHappy new year! I wanted to take a minute to look back at 2025, because honestly it moved faster than any year I can remember in this field. At the start of the year, "agents" mostly meant cool demos that fell apart after five steps. By the end, I was using them every day for real work.
What stood out to me
MCP won. It went from one company's proposal to something every major player supports. In December it was even donated to a new foundation under the Linux Foundation, which is about as clear a signal as you get that it's here to stay. I wrote about why I switched back in June, and I'm glad I did it early.
Coding agents went mainstream. Claude Code, Codex, Jules, Cursor's agent mode, and plenty more. I went from being skeptical to using a coding agent for a large part of my day. I'll write a proper reflection on that at some point, because my workflow changed a lot.
Models got stamina. Early in the year, keeping an agent on track for 20 steps was hard. By the fall, models were running for hours. That changed what's worth automating.
Open weights got serious. Between gpt-oss, the Qwen and DeepSeek releases, and Kimi K2, self-hosting stopped being a big compromise for a lot of tasks. For regulated environments like many I work in, that's huge.
"Context engineering" got a name. Small thing, but it gave people a better way to talk about the actual work.
What I got wrong
I thought visual agent builders would take over more than they did. They found a niche, but most serious agent work I saw stayed in code. I also underestimated how quickly smaller models would get good. By the end of the year, I was routing a lot of work to small, cheap models that would have seemed impossible in January.
What I'm watching in 2026
A few predictions, which I fully expect to look silly by December:
- Evals become a normal part of CI. Not a research thing, a build step. Teams that ship AI without them will have a rough year.
- Agent security becomes the main conversation. As agents get more access to real systems, prompt injection stops being theoretical. I think we'll see some painful public incidents.
- Arabic and other non-English models get much better. There's a lot of investment in the region, and I'm excited to see where it goes.
- Voice agents get good enough for real customer service. Not just demos, actual phone lines.
- Costs keep dropping, so the question becomes "what's worth automating?" rather than "can we afford it?"
A personal note
This blog started as a place to keep my own notes, and it's been really nice hearing from people who found the posts useful. Thanks for reading. If there's a topic you'd like me to dig into this year, reach out. I read everything.
Here's to 2026. Let's build some agents that actually deserve the hype.