Class A · 1:many
Shared agent
One agent, many people. Examples today: Basal, Pen, Pump, Sensor, Lancel. Higher bar — John approves the scope before any code is written. Configure-edits require a reviewer.
Agents · New request
Request a new agent.
Five minutes to file. Pen interviews you, John approves, Pump
scaffolds the Worker, Sensor reviews. You get a working agent in a
few days.
How it works
Six steps from idea to a working agent.
- 01
You file a request
DM @pen in Slack with what you want, or fill out the form below. Pen runs a 5-minute interview to nail down scope, allowed users, and PHI status.
- 02
Pen drafts the persona
Pen writes facets/<name>.md and prompts/<name>-system.md, then opens a PR against Luna_AI_Workforce_Config. The persona is the contract — what the agent is allowed to do, who can talk to it, what it knows.
- 03
John approves
For 1:many shared (Class A) agents, John approves the scope before Pump touches code. For 1:1 owner agents, the owner approves their own. The bar is higher for shared agents.
- 04
Sensor reviews the persona
Sensor adversarially reviews the persona PR against the Cloudflare-Agents anti-pattern checklist. Findings are filed as PR comments. The PR doesn't merge until Sensor's findings are resolved.
- 05
Pump scaffolds the Worker
Once the persona PR merges, Pump scaffolds apps/agent-<name>/ with the right bindings, opens a code PR. Sensor reviews this too.
- 06
John merges; CI deploys
After Sensor approves the code PR, John merges. CI builds and deploys the agent to *.nightluna.com behind CF Access. The new agent shows up on /agents.
The intake template
Copy this into your DM with @pen.
Phase 1 of the platform routes new agent requests through a Slack DM with Pen. Phase 2 wires this form to a Notion DB so the request threads itself. The fields below are exactly what Pen needs to run the interview — pre-filling them shortens the back-and-forth.
Hi @pen, I'd like to request a new agent.
Who's it for: [yourself (1:1) | your team / the company (1:many)]
Owner / allowed users:
- [email protected]
- any other emails who should reach this agent
The job in one paragraph:
[What does this agent do for you? Be concrete.]
Top three use cases:
1. [...]
2. [...]
3. [...]
What it's allowed to act on:
[Read-only? Read/write? Which systems? Slack, Notion, Drive, HubSpot?]
What it must NOT do:
[Things that are off-limits — sending external email, pushing code, etc.]
PHI scope:
[yes / no — does this agent see patient identifiers, trial data, anything you'd treat as PHI clinically?]
Who decides on prompt changes once it's live:
[you alone / a reviewer / John for shared agents — see /governance#change-control]
Class A vs Class B
Two different shapes; two different bars.
Class B · 1:1
Owner agent
One agent, one person. Examples today: Clara, Jonah (Jon's CFO chief of staff). The owner approves their own. Personal overlays — custom instructions, memory off — are self-serve.