The Agent Wars
The next contest in AI is for the work we trust it to do again tomorrow.
On 31 January, I emailed Peter Steinberger about a problem with Clawdbot on Windows. The gateway kept dropping offline. He replied in eight minutes.
That same day, I opened an issue in the project’s GitHub repository. The request was practical: help me keep the agent running. GitHub issue #5549
The project became OpenClaw. In February, Steinberger announced that he was joining OpenAI and that OpenClaw would remain open and independent under a foundation. Steinberger’s announcement
By autumn, several of the largest AI companies were pursuing the same broad idea: an agent that remembers the work, uses other applications and keeps going between conversations.
My January problem now looks like an early version of the industry’s central challenge. Before an agent can take responsibility for a task, the system around it has to be dependable.
The contest for recurring work
On 11 August, SpaceXAI and Cursor introduced Grok Bot in beta: a team of agents sharing a cloud computer, with a coordinating agent assigning work to specialists. Cursor’s acquisition by SpaceX closed three days later. Grok Bot · Cursor
On 8 September, Meta introduced Muse, accessible through its own app and WhatsApp, with a dedicated virtual machine for carrying out tasks. On 29 September, OpenAI announced dots: always-on agents that work across connected apps on their own computers. These are the companies’ product descriptions; sustained reliability has to be demonstrated in use. Meta · OpenAI
OpenClaw offers a self-hosted route. Users can run its open-source gateway on their own computer or server and connect it to messaging applications and AI models. OpenClaw documentation
The designs differ. The commercial ambition is similar: become the system a person returns to when something needs doing.
The view: becoming the default
An agent that knows a project’s history can reduce the work required to explain the next task. Connected accounts, saved preferences and repeatable procedures can make it more useful over time.
That creates a potential advantage. Changing providers may mean rebuilding context and checking permissions again. Portable memory and open standards could reduce that cost. People may also keep several agents and assign each a different role.
The prize is a recurring relationship: research this each week, maintain this website, prepare this report, tell me when something important changes. The provider that earns those assignments gains an opportunity to become the user’s starting point for other services.
The network is already registering more activity. On 30 September, Cloudflare reported that daily requests from AI agents on its network had grown more than 1,700% over a year. That measures requests, not completed jobs or paying customers. It is evidence of activity, not a verdict on commercial success. Cloudflare
The investment question is who earns money when delegated work becomes routine.
