Learn the watch.

Quickstart

Get from zero to a watched fleet in about ten minutes. You'll connect one agent, install one monitoring pack, and watch the command center catch its first anomaly.

1. Create your operations center

Sign up for early access and register your first agent. Each agent gets its own monitoring profile, escalation policy, and audit trail — all enforced out-of-band.

2. Connect a channel

Start with one watched surface — Telegram is the fastest way to see the loop in action. Follow the guided OAuth or bot-token flow:

3. Install a skill

Monitoring packs tell the command center what to watch for. Install drift-watch and send a few test messages through Telegram.

Monitoring loop, explained

The monitoring loop is the entire operating model of the platform. Your agents run wherever they run — the command center watches from outside the execution path, and the loop never stops.

The five phases

1. Monitor. Every agent interaction is observed live — accuracy, tone, performance, cost — across every watched surface. Monitoring is read-only and adds zero latency by design.
2. Detect. Drift, anomalies, and edge cases are caught as they happen: accuracy sliding against its baseline, tone breaking policy, exceptions spiking. Detection runs against the baselines your monitoring packs define.
3. Escalate. When something crosses a threshold, the command center opens an escalation with full context attached — the transcript, the diff from baseline, the policy it tripped. The customer never has to repeat themselves.
4. Govern. Permissions, compliance modes, and kill switches are enforced from outside. A human can acknowledge, resolve, dismiss, or quarantine an agent — or flip its autonomy level — without touching the deployment.
5. Improve. Every human decision feeds back: resolutions tune the baselines, dismissals quiet the noise, feedback sharpens the packs. The loop gets smarter with every intervention.

Why out-of-band

Inline guardrails slow the agent down and break when the agent breaks. The command center watches — it never sits in the request path, never adds latency, and keeps working even when the fleet is on fire. Oversight you can count on is oversight that can't be dragged down with the ship.

Writing an MCP skill

Monitoring packs are how you teach the command center what to watch for. A pack is a small package — a manifest, a set of detectors, and guidance prompts — wired into the monitoring loop.

Anatomy of a skill

• pack.json — the manifest: name, version, permissions, and which monitoring phases may use it.
• tools/ — the functions the agent can call (API clients, parsers, writers).
• prompts/ — guidance the planner reads when deciding how and when to use the skill.

Scaffold and ship

Permissions

Packs declare exactly what they may touch in pack.json. A pack that only reads email cannot send it; a pack without the intervene permission cannot quarantine agents. The oversight plane enforces these boundaries on every check — least privilege is structural, not advisory.

Connectors setup

Connectors make external channels watched surfaces. Each connector handles authentication and streams activity to the monitoring loop — and doubles as an escalation channel when a human needs to be reached.

Google Workspace

Connect via OAuth. Grant Gmail, Calendar, and Drive scopes individually — the command center only monitors what you authorize. Revoking a scope stops monitoring instantly.

Email (IMAP/SMTP)

Any mailbox works: provide host, port, and credentials (or an app password). Every message is monitored, and the same channel carries escalations to your team — subject to your policies.

Telenyx SMS

Provide your Telenyx API key and a sending number. Inbound SMS is monitored like any surface; outbound SMS is the fastest escalation path — urgent exceptions page the on-call in seconds.

Telegram

Create a bot with BotFather, paste the token, and the surface is watched. Telegram doubles as the fastest intervention channel — acknowledge escalations and flip autonomy from chat.

WhatsApp

Connect via the WhatsApp Business API number. Threads stay in context for the oversight plane, so follow-ups days later still carry the full story for any human who picks up the escalation.

Feature flags

Every governance control in the platform ships behind a flag — autonomy levels, escalation thresholds, compliance modes. Flags are enforced by the oversight plane, so changes take effect immediately — no redeploy, no restart, no downtime.

Setting flags

Scoping

Flags scope to global, team, agent, or channel. A flag can be strict for your backend team, supervised for support, and supervised-only-in-Telegram — composed from the most specific scope outward.

Packages

Bundle flags into named packages — starter, supervised, locked-down — and assign a package to an agent instead of managing flags one by one. Packages are how you turn oversight into a shippable product edition.