Mastra co-founder Sam Bhagwat, who previously built the Gatsby front-end framework, has laid out a map of where AI agents are heading — and it ends at OpenClaw. In a recent talk (watch it here), he describes an "agentic spectrum" that runs from raw LLM calls to agents, from agents to harnesses like Claude Code and Codex, and from harnesses to the terminal form: the Claw — a category he names directly after OpenClaw and Hermes Agent.
The Agentic Spectrum
The line between an LLM and an agent, Bhagwat argues, is the loop: tool calls, retries, and persistent state turn single-shot text generation into something you can hand work to. The harness era — kicked off by Claude Code's rise — added planning modes, parallel sub-agents with clean context windows, skills and slash commands, background tasks, context compaction, and the ability to interrupt and steer mid-run instead of taking turns. Cloud harnesses push further: sandboxes, heavy parallelism, PRs pushed straight to GitHub, and always-on operation that keeps working after you close your laptop.
"Steinberger's Law"
The step from harness to Claw adds a heartbeat: cron schedules, notification subscriptions, cloud memory, and a gateway that answers on Slack, Telegram, WhatsApp, or iMessage. The final ingredient is continuous learning — agents that create their own skills and refine them by watching how you work. Bhagwat wraps the trajectory in what he calls Steinberger's Law, a nod to OpenClaw creator Peter Steinberger: any harness that survives keeps absorbing features until it becomes a Claw, because users demand it. Every feature request — can it do more, can it run twenty tasks instead of three, can it work while I sleep — pushes in the same direction.
His closing warning: ordinary people only have mental shelf space for one or two Claws, the same way each phone ends up with one map app and one ride-hailing app. The next battle is not capability but mindshare.
Our Take: The Future Is Claw-Shaped
Hosting managed OpenClaw and Hermes Agent instances every day, we would push the thesis one step further. The agent of the future may or may not be called a Claw, but it will be claw-shaped in the way that matters: you send a request, the bot gets it done, and you get the output. No detailed plan to approve, no step list to babysit. Planning modes and permission prompts are trust scaffolding for an era when agents were not yet reliable enough to be left alone — useful today, invisible tomorrow. The users pulling every harness toward Claw territory are not asking for more oversight; they are asking for less. Delegation, not supervision, is the end state.
If you want to feel what that is like today, OpenClaw Launch deploys a managed OpenClaw or Hermes Agent instance in seconds — with Telegram, WhatsApp, and scheduled tasks built in.