Tommy: a 20-agent OpenClaw framework that runs my day, my research and my health plan

Tommy began as a chief of staff in a Telegram thread. Then the jobs piled up, and one agent doing everything turned out to be the fastest way to get everything done badly. So I split the work the way you would split it across people: one owner per job, clear rules, and a shared place to write things down.

The result is a framework, not an assistant.

The teams

Twenty agents live on OpenClaw, an open-source agent framework, running as a background service on a Mac mini that never sleeps.

• Front door. Tommy, the one I talk to. Requests land in Telegram, Tommy decides who owns them and the answer comes back in the same thread.
• Builders. CodeBot, ContentBot and ResearchBot. They build, draft and dig, and hand the result back to Tommy.
• Personal. Abby, a health coach who knows my plan, and Marcel, who runs the apartment renovation and the device sourcing for it.
• The investment desk. Gordon and thirteen specialists: news, earnings, macro, a red team, charts, value and risk. It got big enough to deserve its own page.
• Ops. A watchdog and a set of scheduled jobs that keep the lights on, back up the config every night and tell me when something stops answering.

What holds it together

One brain. Every agent reads and writes the same Postgres database before and after real work. A decision made on Tuesday in one team is visible to every other team on Wednesday. No agent keeps its own private version of the truth.

Eleven connectors. The agents reach the outside world through MCP servers, the standard plugs that let an AI read and act inside an app. Shopify, Higgsfield for images and video, Metricool for social, Canva, Notion, market data and a few more.

Twenty-nine skills. A skill is a plain-text instruction set an agent loads when the job matches. Deep research, meeting prep, a weekly review, image generation, a stock analyst playbook. When I want a new behaviour, I write the brief the way I would brief a person.

A dispatcher. Before an agent starts a job, a fast model reads the task and decides which brain, how much thinking and which tools it gets. The boring jobs go to cheap models. The hard ones get the expensive one. That one change is why the whole framework runs on a monthly bill smaller than most people’s phone plan.

What broke

All of it, at least once. These are the ones that taught me something:

• Gateway crashes caused by a config file that another process had no business writing to.
• One wrong field in one provider’s settings that took down every agent using it, with error messages that pointed everywhere except the field.
• A sandbox that switched itself off after a container error and left agents running with more access than intended. That became the audit post.
• Too many agents on one expensive model. In September I paused the whole fleet, audited every job and rebuilt it around routing. It came back smaller in cost and bigger in scope.

What I would do differently

Start with the rules, not the capability. Every agent now has a short file that says what it may touch, what needs my yes, and when to stay quiet (nothing between 11pm and 7am unless something is on fire). I wrote those files after the fact. Write them first.