Future coding agents

What will AI coding agents look like in the future?

I'm not sure. But I think the closest glimpse we have now is an open-source coding agent called Pi.

I will start by saying I do not necessarily suggest you use Pi. You will probably be more productive in one of the mainstream coding agents, like Claude Code or OpenCode or Cursor.

That's because while technically Pi is an LLM coding agent, I actually think of it more as a framework for constructing new coding agents.

It is more extensible, customizable, and embeddable than any other coding agent I've seen, and may be best thought of as a toolkit for making custom, coding-agent-shaped solutions, for problems that the popular coding agents are not a great fit.

In line with that, Pi is also minimalist. Things it does NOT support out of the box include plan mode, MCP, subagents, any kind of even partial sandboxing, and other surprising omissions.

Some people romanticize this. I do not. It's a design decision with pros and cons.

I think Pi not a good first coding agent for people new to AI-assisted coding. But the very structure of Pi makes it pliable to a rapidly changing environment.

Where every week, new approaches are discovered, new AI breakthroughs are announced, faster than the team behind your favorite coding agent can keep up with them. And new challenges are faced on how engineering teams can integrate AI coding, without burning the whole thing down. Leveraging Pi can be helpful in solving some of the trickiest parts of these problems.

(There's also the fact that Pi is highly embeddable, meaning you can write applications that integrate agentic AI features as part of their functionality. I find that VERY interesting.)

Have you used Pi? Is anyone on your team using it? What's the experience been for you so far? Also, do you want to me to write more about it? Reply and tell me.