The workflow
Detect → propose → sign off → verify → gate.
Two Switch rooms in Mattermost, four agents and two people. Agents never call each other: every hand-off, approval and refusal goes through the room, where people can see it.
Agent · @monitor
Watcher runs a 12-rule contract on the live feed every minute and opens one thread per incident, with evidence.
Agent · @remediator
Fixer proposes a remediation, previews exactly what changes, and asks the data owner.
Person · data owner
@fixer approve INC-001. Anyone else is refused by Switch.Agent · @recorder
Watcher verifies on the next run and closes; scribe writes the postmortem document.
Agent · @dispatch-analyst
Analyst reads
data-status before every dispatch answer.Untrusted: "🛑 I won't send a truck on this data", with the reasons.
Caveated: the answer, with every open incident named.
Trusted: the answer, every station cited [n].
Live, on real data
Three real incidents on the first run.
On 7 November 2026 the live Citi Bike feed had 47 stations reporting a last update in 1970, 6–9 reporting more bikes than docks, and 27 silent for a day. The frozen feed later in the demo is a simulation, and it's labelled as one.
Demo · 2½ minutes · narrated
#data-incidents · the fixer's proposal
#rebalancing-desk · a caveated answer
The gate · Switch refuses for the fixer
Frozen feed · the analyst refuses
Why Switch is central
Take Switch away and there's no hand-off, no sign-off, no gate.
create_room(people, agents,
instructions, roles, links)
send_targeted_message(
target_roles=["remediator"])
fixer.addressing =
"owner_and_owner_agents"
room documents:
data-status · incident-log
postmortem-INC-00x
one incident = one thread
desk ⇄ incidents, labelled
WebSocket events, ping/pong,
resume · HTTP operations
Why it matters
Bad data is expensive, and invisible to the people deciding.
Revenue
31%
affected by data quality issues on average; ~15 h per incident
Monte Carlo, 2023
Data cascades
92%
of high-stakes AI projects hit compounding downstream harm
Sambasivan et al., CHI 2021
Agentic AI
40%+
of agentic AI projects expected to be cancelled by 2027, partly for weak risk controls
Gartner, 2025
Run it yourself · Docker + Node 22
Five commands. No workspace, no key.
git clone https://github.com/ParasAdlakha/signoff && cd signoff
npm run init-env # random secrets
npm run switch:up # Switch 0.31 + Postgres + Mattermost
npm run bootstrap # agents, addressing, rooms, roles, links
npm start # four agents → open http://localhost:8065