# Daily Vox Brief — 2026-05-25

**Generated:** 2026-05-25 (manual, by Loom — automated curator still not built; second hand-written brief in the series)
**Source:** Today's client-stage dashboard design session, including the Phase 2 agent-integration rework
**Note for Vox:** Today produced a meta-decision worth preserving before it disappears into the spec. Format mirrors the 2026-05-21 brief so this can sit next to the automated ones once they begin.

---

## Lede Candidate

**The "agent-operable from day one" rule — stop building human-only systems if you ever want agents to run them.**

The client-stage dashboard started as a normal internal tool: schema, API, Slack command, cron jobs, manual updates from bookkeepers. Standard CAS-practice fare. Mid-spec, Jimmie asked one question — "can this be updated by my agents?" — and the answer surfaced a design choice that's easy to skip when you're moving fast: does the architecture allow agents to operate this end-to-end later, or only humans? The honest read of the draft was "only humans, with bolt-on automation possible." That's not the same as agent-operable. So the spec went back for a Phase 2 expansion before any code shipped. The rule that fell out of it: every endpoint, every audit-trail field, every permission check is designed so human, cron, and agent callers use the *identical* interface. Phasing is who-calls-the-endpoints over time, not what-endpoints-exist. For the whitepaper's "what makes this team different" section: the cost of building the human-only version first and adding AI later is enormous, paid in retrofits forever. Designing for agents from day one is dramatically cheaper than backfilling — and it changes nothing about Phase 1 except where the audit trail differentiates the actor.

> Raw moment: Jimmie's reply to "can agents update it" was *"yes i want to be able to eventually run most of my company with agents, that is my goal ultimately."* That sentence — delivered mid-spec, before any schema was locked — turned what was going to be a "Phase 2 nice-to-have" into a baked-in architectural requirement. The next two hundred lines of spec are entirely consequences of that one decision moment.

---

## Supporting Beats

- **Permissions-as-data, not code.** When the spec expanded to handle agent writes, the natural temptation was to encode each agent's allowed operations as a switch statement, a config file, or a role enum. The choice instead was to put permissions in a database table — `agent_permissions(actor_id, permission, scope, granted_at, granted_by, revoked_at)`. Adding a new agent becomes a row insert; revoking is a single UPDATE; expanding scope is editable without a deploy. The whitepaper lesson: if you want to scale a team of agents the way a real firm scales a team of people, treat agent permissions the same way HR treats role assignments — as data with an audit trail, not as code that needs engineering effort to change. The corollary is that *who can do what* becomes a question someone in operations can answer and adjust, not a question that requires a developer.

- **The MVP shape survives even when you're designing for AI.** Inside the same session, there was a real temptation to make "advisory delivery" a distinct workflow stage — it's a recognizable thing that happens, advisory clients clearly move through it, why not name it? The decision instead was to start with a flag and promote it to a stage only if lived experience demands it. A stranger might assume agent-operable systems require richer taxonomies up front — they don't. The same "ship the simpler shape, iterate on signal" discipline that applies to human workflows applies to agent workflows. Over-designing the data model for agents is the same trap as over-designing it for humans, and the cost of being wrong is identical: stuck with a state machine that doesn't match the territory.

---

## Skipped (override if you disagree)

- The full 8-stage / 9-flag taxonomy and its transition rules — that's a *finished spec*, not a *transferable lesson*. Vox's readers don't need to know our stages; they need to know how we arrived at the design discipline.
- The ROI math on the dashboard build (~137 hr/yr Phase 1, ~272 hr/yr combined) — useful internally for greenlighting, but spec detail, not story.
- The renumbering of spec sections after Phase 2 was inserted — pure editorial mechanic, with one small craft note: the insert ate a section header on the first try, which is just a reminder that surgical edits to long markdown documents still benefit from a re-read pass before declaring done.
- The specifics of which PKA team members got which permissions in the write-path matrix — internal to our roster, doesn't translate.

---

*Override any of the above by editing this file in place. Vox reads on Mondays.*