Agent Catalog — what the AI does

← Back to AI Cockpit

The inventory of every autonomous capability in the platform: what it does, what wakes it up, how much rope it has, and where you turn it off.

The four families

Family 1 — Chat specialists (the Pulse Copilot)

You type; a router picks exactly one specialist; that specialist answers with its own tools.

Router — picks who answers

• What it does: reads your message and picks one specialist. A fast keyword pass first (bilingual EN/SK); if that is not confident, a small model classifies. A weak pick falls back to General.
• Trigger: every message.
• Autonomy: chooses, never acts.
• Toggle: none — the router is the one part of the Copilot you cannot tune.

Off-domain guard — declines non-work questions

• What it does: refuses questions unrelated to running properties.
• Autonomy: refuses only when highly confident it is off-topic; otherwise it answers normally.
• Toggle: none.
• What it is not: a security control. It filters topics; it does not gate writes, tools, or tenancy.

The specialists

Each is editable at AI Cockpit → Chat Agents — model, thinking budget, tool rounds, response length, plan-then-execute, and the system-prompt mandate. Edits take effect on the next message; no deploy.

⚠️ Know which writes skip the preview. Most Copilot changes show a confirm-with-diff step before committing. Three paths do not: knowledge corrections (edits guest-facing content immediately), template image updates, and an approved guest reply (approving IS the confirmation — it sends). AI-composed guest sends are additionally reviewed by the Outgoing Message Check below.

Coordinator — splits a request across specialists

• What it does: decomposes one request into up to 5 sub-tasks, dispatches each to a specialist, then merges the results.
• Trigger: automatic for multi-part requests.
• Toggle: platform-level; not operator-facing.

Family 2 — Validators (the write/send guards)

The validator list is data, not code — per-tenant rows editable at AI Cockpit → Validators. The tab is the source of truth for which validators run on your tenant. The standard pair:

Each row carries its own confidence threshold and blocking threshold — see Validator thresholds.

The send chokepoint — the thing that actually blocks

A synchronous check sits in front of guest sends; the Outgoing Message Check runs through it. Two deliberate carve-outs worth knowing:

• Access-code carve-out — the guard refuses to block a message carrying a guest's access code. A blocked check-in code is a guest locked out of the building at midnight; that is worse than a bad sentence getting through.
• Access-code delivery guard — after a task is marked done, this backstops the "marked done but the code never actually sent" failure.

Honesty guard (ghost-claim check)

The Copilot checks its own claims: if it says it changed something and no write actually landed, it tells you with a warning banner instead of letting the claim stand. If you ever see that banner, verify the record before assuming the change happened.

Family 3 — Event agents

Maintenance agent

• What it does: a guest message classified as a problem wakes it. It reads context, then either proposes a fix for your approval, or escalates by creating a task for a human.
• Trigger: the maintenance_needed event off an inbound guest message.
• Autonomy: operator_in_loop — it proposes; you approve or reject via WhatsApp buttons. It does not act alone.
• Toggle: AI Cockpit → Event Agents.

Think of it as AI triage-and-hand-off: an unattended guest complaint becomes either a concrete proposal on your phone or a routed task for a human — never a silent fix behind your back.

Sub-behaviours it can propose: access-code recovery, WiFi-details resend, parking-details resend, smart-lock battery triage.

Family 4 — Background agents (the always-on jobs)

The recurring jobs that run your properties while nobody is typing. Each is a named, single-purpose worker over the same trust-gated machinery:

Stay Planner — plans every reservation's task list

• What it does: watches every reservation and spawns its full task plan from your task templates — generate access code, send check-in, schedule cleaning, request review — each at the right offset for that stay.
• Trigger: reservation created / changed / cancelled events.
• Autonomy: creates and cancels tasks; never sends anything itself.
• Toggle: per-rule, in Operations → Task Templates.

Readiness Checker — verifies the real world before acting

• What it does: at each task's fire time, checks the task's conditions against reality (PIN generated? guest replied? cleaner assigned?). All met → completes the task and runs its action. Not met → marks it blocked and pings you on WhatsApp with the exact failing condition.
• Trigger: each task's scheduled fire time.
• Autonomy: completes or fails tasks; a skipped action is never silently "done".
• Toggle: per-template auto-resolve config (Tasks → Auto-resolve).

Outbound Guard — reviews every guest-bound message

• What it does: the send chokepoint from Family 2, worn as a name: reviews every message headed to a guest, blocks or advises per your thresholds, and refuses to block access codes.
• Trigger: every guest send, from any path (auto-send, Copilot, manual).
• Autonomy: can block; you get Refine / Send Anyway / Dismiss.
• Toggle: AI Cockpit → Validators (the Outgoing Message Check row).

Mood Watch — reads how every guest is doing

• What it does: analyses each inbound guest message for language and mood, updates the reservation, and lets mood-dependent rules react (a thank-you after a happy message; extra care after a frustrated one). Reply drafting picks up the detected language automatically.
• Trigger: every inbound guest message.
• Autonomy: analysis only — it writes the signal, other agents act on it.
• Toggle: none (always-on analysis).

Knowledge Keeper — never makes the same mistake twice

• What it does: when you correct an AI answer, writes the correction into the knowledge base so every future reply — from any specialist — uses the corrected fact.
• Trigger: your correction in the Copilot chat.
• Autonomy: writes the correction immediately (this is one of the no-preview paths flagged above).
• Toggle: the Knowledge specialist in AI Cockpit → Chat Agents.

Price Scout — keeps vacancy prices fresh

• What it does: fetches current channel prices for upcoming vacant nights so Pulse's vacancy card always shows what an empty night would earn.
• Trigger: scheduler.
• Autonomy: read-only — fetches, never changes a price.
• Toggle: none (informational).

Current limits

Honest limits of today's autonomy surface:

• Everything acts with you in the loop. The autonomous level is selectable in the UI, but every agent and validator runs operator_in_loop — proposals need your approval.
• One event agent. The Event Agents family currently ships the maintenance agent; more agent types will appear in the same tab as they ship.
• Background agents are named, not separately toggleable. The Family 4 names describe shipped jobs; their controls live in the surfaces listed per entry (task templates, validator rows), not yet in a dedicated per-agent switchboard.
• Self-promotion is not a thing. Agents do not "learn and promote themselves" to more autonomy — autonomy is a setting only you change.

Related

• AI Cockpit overview · Chat Agents editor · Prompts editor · Cost dashboard
• Pulse Copilot user guide
• Validator thresholds