Autonomous ops, on call 24/7.
A standing always-on Claude agent that watches every system in the studio, surfaces anomalies in plain language, and — when it has the authority — fixes them itself.
The problem
The studio runs twenty active businesses on a small team. Twenty status pages. Twenty billing dashboards. Twenty Stripe webhooks, Vercel deploys, Neon branch health checks, and customer-support inboxes. Traditional observability stacks demand you wire a custom dashboard for every property and stare at all of them — which nobody has time to do — or wait for the first angry customer to surface the incident for you.
The right primitive isn't a dashboard. It's an operator. Someone — or something — that has its eyes on every signal, knows what normal looks like for each business, and only interrupts you when something genuinely needs a decision.
What we built
Pulse Agent is a Claude Opus 4.6 process that maintains live MCP connections to fifty-plus tools across the studio — Stripe, Vercel, Neon, Cloudflare, Linear, Slack, Gmail, GitHub, Sentry, Twilio, and so on — plus a thin custom layer that knows which of those signals belong to which business. The agent runs continuously, polls each connector on its own cadence, and keeps a rolling memory of the last twenty-four hours of "normal" for every metric it sees.
When the agent spots a deviation — failed deploy, dropping conversion rate, expired card, support inbox going stale — it writes a structured incident note, scores severity, and either: (a) fixes it itself with the standing playbook (retry the deploy, rotate the key, escalate the ticket), or (b) pings the right human with the full context. No dashboards. No noise. Just the call when you actually need to make one.
It's the first observability tool I've ever owned that I don't check. It just calls me when it needs me, and the call is always worth taking.
How it shipped
Two-week discovery audit to map the integration surface and decide which actions the agent could take unsupervised. Six weeks of build — the MCP connector pool, the per-business memory store on Neon, the severity scorer, and the standing-playbook executor. Four weeks of supervised beta with the studio as customer zero: every action the agent wanted to take was reviewed by a human, and the playbook was rewritten in plain language whenever the agent guessed wrong. By week ten, the false-positive rate was under five percent and we cut the human-in-the-loop step for the lowest-risk action class.
The beta is now open to a small set of design partners. The pricing model: a flat monthly per-tenant plus a metered fee per autonomous action. The math beats a stitched-together PagerDuty + Datadog setup for any team running more than three or four production systems.