A year ago, a personal agent was something you built on a weekend. Today every major AI lab has one, each with its own name, its own personality and its own pitch for why it should be the first thing you message in the morning.

That is the real contest of 2026. The bots are competing for attention: the attention of developers deciding where to build, and the attention of everyday people deciding which assistant gets access to their inbox, calendar and wallet.

It has been a remarkable year to watch. It has also been an exhausting one.

How the race started: OpenClaw

The craze began with a hobby project. Peter Steinberger made the first commit to what became OpenClaw on November 25, 2025. It started life as Clawdbot, a self-hosted assistant you could talk to through WhatsApp, Telegram, Discord, Slack and Teams.

By late January it was the most talked-about open-source project on the internet. It crossed 100,000 GitHub stars by January 30, the same week a trademark request sent it through a quick rename to Moltbot and then OpenClaw. Agents running it even got their own Reddit-style social network, Moltbook, where humans were "welcome to observe."

Less than three months after that first commit, OpenClaw had about 196,000 stars and 600 contributors. Steinberger announced he was joining OpenAI and moving the project to an independent foundation.

OpenClaw proved something the big labs had been circling for years. People want an agent that lives where they already talk, remembers them and does things on their behalf. It also showed how fast that demand could turn into a land grab. Even Microsoft's always-on work agent, Scout, is built on the OpenClaw framework.

The contenders

Hermes (Nous Research). Released in February under the MIT license, Hermes Agent runs on your own server and keeps its memory across restarts. Its signature idea is a learning loop that creates new skills from experience and refines them over time. Hermes is the developer's agent: open, inspectable and built to be shaped. Its August "Herald" release alone included roughly 3,650 commits from more than 650 contributors.

Grok Bot (xAI). Launched in beta on August 11, Grok Bot gives each agent its own cloud computer that signs into the tools you already use. The pitch is a team of "always-on agents" that "finish jobs end to end, and only come back when something needs your approval." Show a bot a workflow once and it saves it as a routine. xAI made it available to Cursor subscribers from day one, a clear bid for developers.

Muse (Meta). Meta launched Muse on September 8 for U.S. users on the web, in its mobile apps and inside WhatsApp chats. It connects to email, calendars, payments, fitness apps and shopping, and it keeps working after you close the app. Muse is free to start, with $20 and $100 monthly tiers. Meta's reach makes Muse the consumer front-runner. Its privacy history makes trust the central question for adoption.

dots (OpenAI). Three weeks later at DevDay, OpenAI introduced dots, "always-on agents built to handle everything." You start with one primary dot and give it a name, and OpenAI imagines teams of dots working on your behalf in the background with minimal oversight. It launched for ChatGPT Pro and Business Premium users, reachable through Slack, Teams and other workplace tools.

Claude Duet (Anthropic). The newest entrant arrived in late October as a waitlisted research preview for U.S. Pro and Max subscribers. Duet sends a morning brief, works in the background across connected apps and asks before sensitive actions like sending messages or paying. It carries over the memory, skills and connected apps people already use in Claude, a bet on continuity. CBS News framed it as Anthropic's answer to Muse and dots.

And the long tail. Add Microsoft Scout, a growing list of OpenClaw forks and agent features inside nearly every productivity app, and the field gets crowded fast.

Two audiences, one scarce resource

Look closely and every launch is aimed at one of two audiences.

Developers are courted with openness and leverage. Star counts, contributor numbers, MIT licenses and coding-tool bundles all say "build here." A developer's attention compounds, because the skills, plugins and integrations they write make the platform more useful for everyone else.

Consumers are courted with presence and personality. Muse lives in WhatsApp. Dots get names and bubbly avatars. Duet greets you with a morning brief. Each product wants to become the habit, the one thread you open before any other.

Both audiences are being asked for the same scarce thing. A personal agent only becomes useful after you give it your context, your accounts and your corrections. Nobody has enough of that attention to spread across six agents.

The temptation to collect them all

It is easy to become a strange kind of agentic Pokémon trainer. You install OpenClaw on a spare machine, try Hermes for its memory, join the Duet waitlist, sign up for Muse because it is free and keep a dot around for work.

For a few weeks, that feels like progress. Every agent does something impressive in a demo.

Then the costs show up. Your preferences live in five places. Each agent knows a little about you and a lot about nothing. You spend evenings re-explaining the same context, reconnecting the same accounts and comparing outputs that all started from scratch. The novelty wears off before any one of them has had the chance to get good.

The name of the game: pick one and hill-climb

In July I wrote about how organizations climb the hill of their own outcomes, building evaluation loops that make every cycle of work a little better than the last. The same idea applies to one person and one agent.

In my experience, the agents that feel genuinely useful are the ones their owners have been shaping for months. Their value comes from accumulated context: corrections, standing rules, saved routines and a clear sense of what the person cares about. That accumulation belongs to you, and it is the part no launch event can hand you.

How to choose

Start with the outcome you want, then match the tool to it.

  • Where does your life already happen? If your day runs through WhatsApp, Muse has a head start. If it runs through Slack and Teams, dots or Duet fit naturally. If you want full control on your own hardware, look at OpenClaw or Hermes.
  • How much do you want to own? Self-hosted agents give you inspectable memory and skills, and they ask for your time in return. Hosted agents take care of infrastructure and security, and they ask for your trust.
  • How much should it do without you? Some agents are built to run with minimal oversight. Others stop and ask before they send or pay. Choose the one whose approval model matches your comfort today.

How to climb

Once you have chosen, put in the work.

  • Teach it once, in writing. Give it standing rules in plain language: who matters, what to escalate and what it should never do.
  • Correct it every time. Each edit and rejection teaches it something. A bad draft you quietly ignore teaches it nothing.
  • Widen its permissions slowly. Start with drafting and preparation. Let it act for you only after it has earned that on smaller tasks.
  • Turn repeats into routines. When you ask for the same thing three times, make it a skill or a scheduled job.
  • Review what it did. A weekly look at its actions shows where the boundaries need to move.

Who wins the bot wars

The labs will keep shipping. By this time next year there will be new names on this list, and today's headline features will be table stakes. Switching will always be one sign-up away.

The people who come out ahead will be the ones who chose deliberately and invested patiently. They will have agents that know their priorities, respect their boundaries and get a little better every week.

Pick the tool that fits your life. Then put in the work to make it yours.