I keep opening with “in the world post-OpenClaw,” but it’s important.
That was a flash point. That was where the public started to understand what some of us have been talking about for the last five years.
Here goes,
In the world post-OpenClaw, I’m having a harder time taking new agent frameworks seriously unless they’re long horizon solvers.
We have task runners now. We have orchestration layers, tool routers, workflow graphs, prompt chains, retry loops, and every flavor of “the model calls the thing.”
The boat has sailed. Somebody brought a fog machine and jello shots, and well… you know.
Our niche went from something only a few deeply talented people were working on seriously, to something everybody thinks they can do.
I’m not going to fault everybody. Everybody is fine with me. I like everybody. Mostly.
But it is a sign that those of us who have been pushing the technology since the beginning… need to push a little harder.
Own the continuity
So where do we go from here?
At the dev-facing prosumer level, the big unsolved problem is not whether an agent can perform a task, but whether an agent can stay oriented while the work stretches across hours, days, or weeks.
It’s easy for someone like me to look at this and scream, “Yeah, that’s solved. I built one for so-and-so last month.” I think a lot of us who have been at this for a while feel that way.
But there are still no mass-market solutions that address this in a meaningful way.
The next generation of general-purpose agents needs to think through this carefully. Simple memory is not enough for an agent to hold the shape of a project over long periods of time.
It has to remember why a decision was made without turning every prior message into wet cardboard. It has to hand work from one agent to another without losing the constraints, the invariants, or the reason the work mattered in the first place.
That’s one place where things got interesting in 2026.
Bring your agent, take the tool
Another is challenging old assumptions.
In the past, the assumption was that you subscribed to a specialized agent system of some kind. That agent did something, and then it went home.
Agents were specialists, and you needed a few of them, maybe a dozen of them, to handle your AI workloads.
What’s exciting is how that’s changing.
Today, you have an agent you like. That agent ties into other tools, some of which may have agents of their own. But you know what they don't need?
A dedicated inference stack.
From the standpoint of a software engineer building for AI workflows, this is monumental. It means you no longer need to hire an inference provider for every offering you put out.
The formula gets much simpler:
Bring your agent.
Take this pretty little tool.
Plug that baby in.
You know what else that means?
You can get experimental again.
You can let apps become a kind of magic again.
All of a sudden, AI apps become really cheap to develop.
And the sky opens back up for everybody, who I’m totally okay with.
Mostly.