FIELD NOTE 11 / AGENTS IN A BUSINESS
Give every agent a job description.
Build a small, accountable agent team before multiplying the headcount.
An agent count says little about what a business can trust. Give each agent a bounded job, a responsible human, and a visible outcome. Expand after the first role works.
What it looks like
Five general-purpose agents, all with the same credentials and shared memory.
Researcher: read approved sources → draft with citations. Reviewer: inspect draft → findings only. Human owner: approve or reject. Publisher: approved artifact only. Each role: its own permitted tools and data.
Put it to work
- 01
Begin with one role and one owner.
Write a one-page job description: inputs, allowed tools, prohibited actions, output, success check, escalation, and shutdown. Pilot it on a small, representative set of real tasks. Record failure cases as carefully as successes.
- 02
Separate identity, memory, and authority.
Hermes profiles and OpenClaw multi-agent routing support separate agent state. Confirm the behavior of the installed version. Separate folders or personas are not a security boundary by themselves. Use actual account permissions, tool restrictions, and appropriate process or container isolation for data that must remain separate.
- 03
Stage the handoff into real work.
Keep research, review, and publishing distinct. A result should carry its source, date, checks, and approval state. Test an expired credential, a duplicate run, a failed check, and the stop control. Then add the next role when it solves a specific bottleneck.
Give this to your AI
Adapt the brackets to your task.Design a small agent pilot for [business workflow] using [Hermes or OpenClaw]. Read the current official documentation and identify the installed version before giving commands. Define one agent role, one human owner, approved data sources, account permissions, memory boundaries, allowed tools, output, verification gate, retry and run limits, escalation, and shutdown. Include a read-only trial and failure tests. Keep deployment as a proposal until the pilot design is approved.
The 100+ figure covers agent deployments across two years, including OpenClaw, Hermes, Paperclip, and GrokBot. It is different from a subagent dispatch count and does not mean 100 agents are currently active. This guide does not expose any private business fleet configuration.
Add a role when it removes a real bottleneck, then make it accountable.
Sources & further reading
Personal counts are from Paul’s prompt-history analysis. Tool guidance was checked against these sources in September 2026. Confirm current behavior in your installed version.