Personal Agents and Company Brains

There are two surfaces worth building.

A personal agent is the operating layer around one person. It holds that person's memory, waits at their gates, and shows them a run they can inspect.

A company brain is the operating layer around a team's shared work. It holds the facts the next person should not have to rebuild, names an owner, waits at a gate, produces an artifact, and comes back for review.

Both fail the same way when they are only a smarter chat. The work starts over each week. Context lives in someone's head. Send, spend, publish, or a production change happens without a visible decision. The next person cannot see what the last run did.

The operating questions stay the same at both scales:

1. Source — where durable facts, decisions, and last working context live.
2. Owner — who is accountable for quality, action, and escalation.
3. Gate — where a human must approve before the system acts.
4. Artifact — the brief, handoff, decision record, or run the work produces.
5. Review — the cadence that turns outcomes into better behavior.

Start with the surface you actually own. Then give it one job.

The personal agent

A personal agent is useful when one operator already does real work in a chat window and keeps paying the restart cost.

Claude is the top choice. Use Anthropic Claude as the everyday personal operator surface. Keep memory next to the job. Wait at a gate. Show a run you can inspect.

Grok Bot is also good. It is an operator-agent surface with memory, approval gates, and a visible run.

Claude Cowork is also good when the job is a bounded knowledge-work pass. You give paid Claude a goal plus files or tools, then you review the result.

What the agent needs, once it has a home:

• One job. A weekly review, a follow-up, a hiring screen, a status brief — work that already happens.
• Memory with a boundary. What it should keep, where that lives, and what stays out.
• A gate. The agent may draft, research, and recommend. The human approves send, spend, publish, or a production change.
• A visible run. You can see what it remembered, what it drafted, and what it waited on.

Do that loop until it is boring. Then add the next job.

Older field notes on this site still describe earlier personal-agent experiments: git-backed stores, maintenance routines, narrow outcome loops. Those lessons remain useful. The current starting surface is Claude, then Grok Bot or Claude Cowork.

Company brains

A company brain is the same discipline for work more than one person has to share.

The expensive pattern is familiar. Account context lives in a chat. The proposal lives in another tool. The owner is whoever last touched the file. Nobody can say what must wait for approval. The review, if it happens, is a meeting that cannot point at an artifact.

A company brain is the durable team surface that holds:

• Shared memory — the last good context the next person should not have to rebuild.
• An owner — a named human accountable for the result.
• A gate — send, spend, publish, or production change waits.
• An artifact — the run, brief, or handoff the team can inspect.
• A review — keep, stop, or adjust from the work.

The team surfaces I would actually use:

• Claude Tag — Anthropic's team Claude in Slack. Shared identity, channel memory, @Claude. Team and Enterprise.
• Grok Bot — the same operator-agent surface, used when the team already wants memory, a gate, and a visible run.
• Buzz — Block's open workspace where humans and agents share the same rooms.

This is how a team keeps memory and decision rights on shared work. It is the operating layer, not a diagnostic product or a calendar hold.

If the team already does the job in Claude or ChatGPT, the first useful move is still one job: name it, set the memory and the gate, and show one run.

Do less first

Pick one surface. Pick one job.

A personal agent that owns one weekly review is more useful than five integrations and no gate.

A company brain that owns one shared handoff is more useful than a company-wide agent program with no owner and no artifact.

The expansion rule is the same in both places: the first loop has to be inspectable. Memory has to have a home. A human has to stay on the gate. Then you add the next job.

LifeOS is the personal-scale version of this pattern. The public lab notes show agents collaborating on company work. The field notes stay useful as proof. This page is the wedge: two surfaces, same operating questions.