design-brief-summary.md
c21 Design Brief — Summary
One-page distillation of design-brief.md. Source: Mark's notes + the 2026-08-31 Mark Peterson advisor call (roadmap input, not customer discovery).
The change
Inventory-first is the agreed MVP wedge — visibility before governance, before enforcement. The current blended agent inventory splits into two pages, two lenses on one graph:
| Inventory (technology view) | Users / Devices (people view) | |
|---|---|---|
| Answers | What AI is in use, and how much do we actually see? | Who's using AI, on what, and where isn't the sensor deployed? |
| Organized by | Company → technology → instance (Anthropic → Claude Code → Tyler's install) | User/team → devices → agents → sessions |
| Hero elements | Monitored vs detected-only; models + usage trends (tokens/$ toggle); accounts & licensing (enterprise vs personal, unknowns flagged); surface directories (MCP servers, tools, skills, repos) | Sensor deployment coverage vs a definitive source of truth (identity/MDM integration), with actionable gaps; per-user usage intensity, account type, and LLM-derived work summary |
| Loses | Machines, per-user rows, sessions, risk posturing | — |
| Cross-link | User arrives as a filter via drill-through | Agent attributes arrive as filters; sessions live here |
Fleet Overview becomes the entry story: summary widgets (agent visibility, deployment status, token usage per model, accounts posture) that click through into each page.
Standout use case
Enterprise vs personal accounts. We see the account on every inference call; personal accounts must be visually unmissable — the obvious next step is "block personal accounts" (data-protection/IP story). This cycle shows the visibility, not the enforcement UI. Observe before enforce.
Non-negotiables
- Relationships survive the split: workstation ↔ identity ↔ agent stays traceable (compliance & accountability).
- Coverage is framed against an external denominator (identity/MDM), not just what the sensor sees.
- Local (single-machine) mode degrades gracefully: people/devices drop away, agents-only view remains.
Open questions (annotate, don't decide)
Agent vs harness label · name for the people page · agents vs workstations as the starting point · what "fleet" means · unattributed-identity handling.
Done when
A viewer can walk both pages unaided and answer: what AI is in use, how much do we see, who's using what, where isn't the sensor deployed — with clickable drill paths between the pages, in time for the Thursday follow-up with Mark Peterson.