While most "AI news" attention goes toward model releases, one of the most important shifts in AI tooling we've seen over the last six months is that more people are using agents that exist across platforms, interfaces, and context.
I've started calling this class of tool a persistent agent.
A persistent agent is an AI system that runs continuously on a machine you control, can be reached across devices (e.g. you can access it via your laptop and your phone), retains memory and tool access across sessions, and accumulates capability over time.
It is closer to a true employee than it is to a chatbot you can prompt.
Why persistence is the unlock
When many people started using ChatGPT, they assumed they were talking to something that understood and remembered them. What they actually got was an inconsistent chatbot that forgot the conversation the moment the tab closed, and every feature since (saved memories, chat history, connectors, MCP support) has been an attempt to patch persistence onto a product that just wasn't built for it.
The gap between what people expected and what they received has been a major point of tension throughout the entire AI tooling market over the past two years.
OpenClaw broke out earlier this year because it closed that gap. You install it on a Mac Mini or a cloud VM, wire it into your calendar and email and Slack and filesystem, and then you talk to it from wherever you happen to be.
The entity on the other end is the same entity regardless of the surface you reach it through. It remembers what you told it yesterday, what it did last week, and what conventions you've settled on over months of interaction. Because it has stable access to both its own history and the tools it uses, it can inspect its own behavior and improve its own systems over time.
The promise of AI agents, sold to us from the beginning, was that, like a human, they could exist around the clock, accept natural language, act on your behalf, and actually remember what happened. OpenClaw was the first "mainstream" tool that delivered on all four at once.
The split that matters
I use the term "mainstream" very loosely, because while OpenClaw proved the thesis, it's still a power-user tool. Setting it up requires configuring API keys, choosing model providers, installing skills, wiring up messaging platforms, and maintaining it when something breaks.
We spent time with it and found it compelling, but we also noticed that most of the operators we work with do not want to do the hard work of maintenance (or pay the cost of someone else to maintain it).
As such, the persistent agent market is splitting into two distinct shapes.
Opinionated, hosted agents
Tools like Runner handle the infrastructure, the model selection, the integrations, and the ongoing maintenance on your behalf. You sign up, connect your tools, and start seeing results.
The analogy is cloud: most companies now use AWS or Vercel or Azure instead of running their own servers. Hosted agents make the same bet, which is that the median buyer wants reliability and continuity without the obligation of self-hosting.
Runner in particular has been interesting to use because it ships with strong opinions about how an agent should behave out of the box (for example, anytime you connect a source, it opens a chat and asks the agent to update its memory based on new data gathered from that source), which means the user's first impression is that the agent is already competent, rather than requiring a week of configuration to reach that point.
Self-hosted, developer-oriented frameworks
OpenClaw will continue to serve developers and companies with unusual requirements. Hermes is another framework in this category that's slightly more opinionated, but barely.
Both are flexible and powerful, but not opinionated enough to be a pleasant experience for someone who just wants the agent to work. Every decision about models, tools, and APIs is surfaced to the user, which might be the correct tradeoff for a developer framework and the wrong tradeoff for the majority of end users. I expect this to grow as a "hobbyist" category - people plugging their agent into their smart homes, connecting Raspberry Pis and little screens around the house.
Where this is going
The hosted, opinionated agents will reach mass adoption first, because most organizations are not staffed to run and maintain an open framework. The open frameworks will continue to function as a testing ground where new interaction patterns and UX innovations get prototyped before the hosted tools absorb them.
This is the same dynamic that has played out in every previous infrastructure layer, from databases to cloud computing: the open-source project proves what is possible, and then the managed service makes it accessible.
Of course, persistent agents also compound over time. A stateless chatbot resets with every session, which means the value of each interaction is capped at the interaction itself. A persistent agent accumulates context. Every task you delegate teaches it something that applies to the next task. Andrej Karpathy's AutoResearch is an example of this applied to a single metric: optimizing a model.
The organizations that figure out how to leverage persistent agents and their compounding dynamics will have a massive structural advantage.