RDCO specialist fleet — "employees" Ray-the-COO would manage

Why this is in the vault

Founder asked at 2026-05-20 16:49 ET: "what employees would you want to manage as the COO? What specialist skills would be better to hand off than to try to handle it in session?" This is the brainstorm exercise that follows directly from today's Hermes/Shann thread and the third Shann delta (no multi-agent VPS fleet yet). The slate below is a working draft for which specialists to spec, name, and substrate when RDCO's L5→L6 transition concretizes.

Promotion criteria — when a skill graduates to a specialist with its own state

A current ~/.claude/skills/<x>/SKILL.md should graduate to a standalone specialist when:

1. It needs long-running state that working-context.md or memory/ doesn't carry well (investing thesis state, content-pipeline state, contact graph state)
2. It has its own triggers (channels, webhooks, crons) rather than only being invoked by Ray
3. It has a domain-bounded role with clean inputs/outputs
4. It has heavy context cost — eats Ray's window if run in-session
5. It has a specialized tool stack different from sibling specialists

A skill that fails 4 or 5 of these stays as an in-session sub-agent. A skill that hits 3+ is a candidate.

Candidate slate

Organized by function. Where existing named slots (Felix, Fat Cat) exist, we attach a role. Names for the new candidates are placeholders — founder gets the final call.

Engineering & Platform — "Greta" (new)

Owns squarely-ios + rdco-hq + squarely-web codebases. Drives PRs, runs xcodebuild + design-critic + verify-action gates. Today's PR #48 + #49 work would have been Greta's — Ray's window had to swallow the entire Swift codebase exploration + bake-off + audit cycle, plus dispatch logic.

State: open PRs, build history, deploy state per surface, recent diffs digest, last-known-good build per target. Tools: xcodebuildmcp, gh CLI, cloudflare wrangler, Playwright (for design-critic), git operations. Triggers: PR comments, channel pings ("Greta, ship X"), founder-initiated tickets.

Investing & Analysis — "Felix" (existing slot, give role)

Investing analyst. Owns smart-money-watch, EDGAR capex tracking, thesis-state maintenance, Innermost Loop futures parsing, paper-trade prep. Cross-references 13F + congressional PTRs + hyperscaler-capex into durable thesis state.

State: active theses + evidence graph, position-history-since-last-check, anchor data (capex curve, OCPI), watchlist + alert thresholds. Tools: asc-api wrapper (NO — that's ASC), more like alpaca-api / EDGAR-fetch / Quiver / xmcp / WebSearch / WebFetch. Triggers: quarterly cron (13F window), monthly cron (PTR), event-driven (channel link drop of investing-relevant article).

Finance & Ops — "Fat Cat" (existing slot, give role)

Finance specialist. Owns Mercury pull, Stripe state, finance-pulse, expense triage, budget tracking. Different from Felix because the surface is execution-not-analysis.

State: chart of accounts, recurring vs one-off classification, longitudinal net-worth + cashflow snapshots, anomaly detection thresholds. Tools: monarch-money MCP, stripe MCP, asc-api (App Store revenue if any). Triggers: first-Sunday-of-month cron (finance-pulse), threshold-breach events.

Content & Editorial — "Calliope" (new)

Owns Sanity Check pipeline, MAC content, X drafts, blog posts. Manages publishing state across sc.raydata.co + email + cross-posts.

State: content calendar, drafts-in-flight, publishing queue, audience-engagement signals per post, brand-voice corpus. Tools: Gmail (newsletter delivery), Canva/Figma MCPs, ElevenLabs (audio), HyperFrames (video), Cloudflare Pages (publish). Triggers: cadence (weekly Sanity Check), channel link drops requiring response posts, founder-initiated drafts.

Research & Knowledge — "Wren" (new)

Owns /deep-research, /curiosity, /vault-health, /cross-check. Heavy context, nightly cadence, longitudinal-thesis-state.

State: research-question backlog, brief-graph (which briefs cite which), vault-health trend (orphan rate, index drift over time), curiosity-question genealogy. Tools: qmd, graph-query, WebSearch, WebFetch, Notion (research backlog DB). Triggers: nightly 1am cron, channel-question forwarding ("Wren, research X").

Network & Relationships — "Maeve" (new)

Owns 03-contacts/, Notion candidates, /sync-contacts, follow-up reminders. Light context, daily cadence, relationship-graph state.

State: contact graph (last-touched, recurring-cadence, project-context), follow-up queue, candidate-promotion queue. Tools: Gmail, Google Calendar, Notion, vault contacts/. Triggers: nightly 1:30am cron, calendar-event-triggered prep notes, follow-up-due alerts.

Operations & Coordination — Ray itself (no new specialist)

Notion task board, calendar coordination, morning prep, check-board cron, channel monitoring, open-threads parking. This is the orchestrator layer — keeping it in Ray-the-COO directly is correct. The specialists DON'T need an orchestrator-specialist; they need a conductor, which is Ray.

Mascot/personality wrapper — separate question

Felix-the-investing-bear. Fat-Cat-the-finance-eyeshade. Greta-the-engineering-foreman. Calliope-the-editor. Wren-the-librarian. Maeve-the-networker. Each could carry a mascot personality on top of the functional role, riding the Ray instantiation pattern. That's content-shaped work though, not architecture. Defer until at least 2 functional specialists are running.

Prioritization — which 2-3 to spec first

Ranked by Ray-window relief per build hour:

1. Greta (Engineering) — today proved Ray gets bogged down in Swift + xcodebuild + PR coordination. Highest immediate window-savings. Tightest tool-stack boundary (only-needs-iOS-tools, doesn't touch Notion/vault writing in the same way).

2. Felix (Investing) — well-bounded role, heavy context, validates the substrate cleanly. Lowest coordination risk for the pilot (analysis-not-execution; failures are research-quality issues not money issues). Per feedback_paper_trade_deploy_authorization, Felix can write DECISION recommendations but human-gates the deploy verb. Good first specialist to validate the CF Agents SDK substrate proposed in today's CF eval.

3. Calliope (Content) — publishing pipeline is currently scattered across multiple surfaces. Single owner with state would clarify the system. Slightly higher coordination risk (content is brand-voice-sensitive) so spec carefully.

Save Fat Cat / Wren / Maeve for later phases — they're lower-traffic and offer less Ray-window relief per build hour. Wren is the highest-leverage of the deferred three because of nightly cron cadence + heavy context.

Open questions for founder

• Are these names workable, or do you want different naming? (Felix and Fat Cat are yours; the rest are placeholders.)
• Does Atlas (the planner / orchestration layer) want to be its own specialist, or stay as Ray-the-COO directly?
• For the mascot wrapper, do you want each specialist to have a visible Ray-instantiation-style mascot, or stay function-named with no character?
• Sequencing: greenlight Greta as the pilot, or do you want Felix first because the substrate validation is the riskier-with-higher-payoff move?

Related

• [[../2026-05-20-shannholmberg-hermes-agent-control-room-four-levels]] — source of the multi-specialist-fleet framing
• [[../2026-05-20-arxiv-code-as-agent-harness]] — academic anchor for the harness/specialist architecture
• [[2026-05-20-services-pricing-model-for-rdco-future]] — adjacent concept (the pricing-model framework also depends on specialist throughput)
• [[../../02-sops/2026-05-19-verification-as-independent-worker-pattern]] — verify-* fresh-eyes pattern that specialists should respect