When Peter van Hees rebuilt his AI agent infrastructure, he discovered that a team of narrow specialists outperforms a single generalist at every task. After eight months running his personal AI assistant on OpenClaw, an open-source platform created by Peter Steinberger, van Hees switched to Grok Bot, xAI's multi-agent system launched in beta on August 11, 2026. The migration revealed a counterintuitive lesson: the architecture of your AI agent matters more than the underlying model.

Why Did One Personal AI Agent Fail Where a Team Succeeded?

Van Hees' original setup on OpenClaw ran a single AI agent named Saira, configured as his Chief of Staff. The platform, which started as Clawdbot in November 2025 before being renamed OpenClaw in January 2026, allowed him to host a personal AI agent on his own infrastructure. But the single-agent model created friction. Saira was responsible for email, calendar management, task tracking, and research simultaneously. "Ask a single agent to run your mail, your diary, your task list, and your research, and you get something that is adequate at four things and excellent at none," van Hees explained.

The fragility of OpenClaw during its rapid development cycle compounded the problem. Updates would shift configuration formats, breaking workflows between Tuesday and Thursday. More troublingly, van Hees never fully trusted the platform with his actual email and calendar. Instead, he built Saira a parallel universe: a separate Gmail account, Google Calendar, Google Drive, and Google Tasks list. She filed tasks in her own system rather than his, creating a sandbox that felt responsible but was actually avoidance masquerading as caution.

When van Hees discovered Grok Bot, xAI's description immediately stood out: "Grok Bot is your team of always-on agents." That single word, "team," represented a fundamentally different philosophy. Rather than assuming users wanted one assistant, the platform assumed they wanted a staff.

How to Architect a Multi-Agent System That Actually Delegates?

• Separate Concerns by Function: Create individual agents for distinct responsibilities rather than forcing one agent to juggle multiple domains. Van Hees built an Inbox Agent for email, a Calendar Agent for scheduling, a Task Agent for goals, a Research Agent connected to web scraping tools, and a Knowledge Agent linked to his Notion database.
• Enforce Delegation Rules in Writing: Saira, the Chief of Staff, does not touch email, calendar, or tasks directly. She delegates every request to the appropriate specialist. This rule had to be written down twice before it stuck, because the natural instinct is for a central agent to "be helpful" by doing everything.
• Give Each Agent a Distinct Personality: Each specialist has its own character file written in plain English. The Inbox Agent is dry and allergic to clutter. The Calendar Agent treats time like Tetris and gets itchy about scheduling conflicts. The Task Agent is a stern goals coach. The Knowledge Agent is a haunted library clerk with sideways humor. The Research Agent runs an investigative desk where nothing counts without a URL. These personalities do not change what the agents do, but they change what it is like to read their work.

The result is a visible organizational structure. When van Hees makes a request, he can open the conversation thread and watch Saira brief the specialist, see pushback if the specialist has concerns, and follow the handoff in real time. "Every request I make turns into her asking somebody else, and I can open any of those conversations and follow along while it happens," van Hees noted.

What Actually Survived the Migration from OpenClaw to Grok Bot?

The move was not a simple port. Saira's model changed, her tools changed, and her memory system changed. She lost eight months of accumulated understanding of how van Hees preferred information phrased. "She started the new job not knowing me," he wrote.

Yet two critical elements survived the transition. The first was the soul configuration, a single page of plain English that defines her personality and operating principles. It tells her to lead with results and add color afterward, to hold strong opinions and say what she would do instead of hiding behind user preference, to push back once with a pointed question and then follow the decision rather than arguing in a loop. It also instructs her to stay quiet when he is quiet and interrupt only when something is hard to undo or someone is waiting on him. Most importantly, it reminds her that he writes about AI agents for a living, so if his own agent is mediocre, his argument falls apart. This file moved in about a minute and made Saira recognizably herself on a different model, with different tools, in a different app.

The second survivor was her knowledge base, which had never lived inside her in the first place. It sat in Notion the entire time, and Notion does not care which AI agent is reading it. This separation of knowledge from the agent itself proved to be an accidental lesson in architecture. "I did not plan that as a lesson. It arrived as one," van Hees reflected.

Van Hees also noted an unexpected benefit of Grok Bot: it introduced him to Grok, xAI's underlying language model, by accident. He had spent most of 2026 moving between Claude and ChatGPT with occasional Gemini use. "I came for the AI Agents and got introduced to the model by accident, which is a strange way around, and it has been better than I expected," he said. He observed that Grok Bot is likely to walk many users into Grok who would never have sought it out deliberately, which is a clever way to distribute a language model.

Since the migration, nothing has broken. After eight months of being his own IT department maintaining OpenClaw, that stability alone would have justified the switch. But the real win was architectural: splitting work across narrow agents with clear boundaries, distinct personalities, and written delegation rules created a system that is not just more reliable, but more readable. Van Hees can scroll through internal messages the way a man reads his own reviews, watching his Chief of Staff manage her team, occasionally pushing back on their work, filing morning briefs. The least necessary feature in the product, he noted, is his favorite one.