Snowden Zeng

Portfolio Ops Room — run the portfolio from one Switch deployment

Flip the Switch Hackathon/Nov 2026

This room exists because deadlines kept getting lost. Program 67's deadline system was built after prepared entries died waiting on decisions nobody saw in time — and tonight this entry nearly died the same way: its close was recorded in our register as 13:29 ET; the live form was still open hours later, showing a close of ~23:29 ET. What caught it is the board below: during the live demo, the agent's own deadline list carried this hackathon's corrected deadline. A solo founder runs a portfolio of ventures (Program 67) with a fleet of AI agents. The portfolio's state lives in three places — a deadline register (123 tracked commitments, 97 of them active), a watch layer monitoring the machines, and GitHub — so every status check meant leaving the conversation and opening another tool. The Portfolio Ops Room puts a staff agent inside the room where the team already works. Ask @ops-agent "deadlines", "gates", "money", "health", "brief", "status", or "prs" in the #portfolio-ops Mattermost channel on our self-hosted Switch deployment, and it answers in ~1–2 seconds from real data: the live deadline board (overdue, due ≤24h, next five), the founder actions waiting, overdue payments, per-monitor watch health, and open pull requests — with the data's snapshot time printed on every answer, because a snapshot should never pretend to be a live query. A second agent, the Watch Agent, posts in the room when a fleet monitor flips (its flip logic is dry-run verified; no monitor flipped tonight), and a portfolio digest lands twice a day on a timer. By the numbers: 7 questions · answers in ~1–2 s · the ops snapshot refreshes every 5 minutes and every answer prints its snapshot time · 123 commitments · 1 operator · during the demo the board carried real overdue money next to a genuinely DOWN monitor, both reported as-is. Built for solo founders and small teams running many projects with agent help: the board moves into the conversation, one mention away. No invented numbers — the agent reports the real board, including the uncomfortable parts. --- ◇ **Track 02 — the Creator Room (a second room, a second agent).** The same Switch deployment hosts a second room for the other half of the operator's working life: a media pipeline. Its Producer Agent — deterministic, like the Ops Agent, and registered with Switch the same way — answers eleven questions (`ideas`, `script`, `renders`, `clips`, `next`, `status`, `credits`, `review`, `sources`, `health`, `help`) in its channel from a labelled snapshot of the real pipeline files, and it publishes that snapshot — plus the pipeline state and its review gate — as two room documents, so the documents and the answers move together when the snapshot refreshes. The room's discipline is the point: the producer reports state only — it does not generate, post, or clip — and it degrades honestly (last-good snapshot, labelled, or an explicit "unavailable"), never an invented number. The pipeline it reports on is dormant: the clipping lane was built and proven on samples in October, has never posted, and has earned $0; exactly two MiniMax H3 test renders exist; and generation is blocked (2 of 150 free credits left; the cheapest clip costs 28) on a production-route decision only the founder can make. A personal workflow, run the same way as the business one: the creative work's true state, one mention away — including the parts that are stuck. Blog post (Best Blog track): "Two Rooms, One Operator" — https://dev.to/snedz/two-rooms-one-operator-no9