AI Agents automation

Overview

Every specialised AI Advisor can run unattended on a nightly schedule, in addition to the interactive on-demand path. The automated batch selects eligible entities for a tenant, invokes the relevant advisor (typically in act mode) against each, and records the outcome — all scoped by per-tenant policy that you control.

Automation is configured from Artificial Intelligence > Manage AI Agents automation and observed from Artificial Intelligence > Review Scheduled AI Advisor activity.

Important

Automation is disabled by default for every tenant. Nothing runs unattended until you explicitly enable the ML Advisor and/or the Components Advisor for a tenant and pin them to a provider. The decommission guard (the ability to disable entities automatically) is also off by default.

Note

The AI Concierge has no automated path — it is interactive-only by design. Automation applies to the five specialised advisors only.

Managing automation per tenant

The Manage AI Agents automation page lists every tenant with its automation state. Two families of automation are configured independently:

• ML Advisor — the model-inspection advisor, with its own enable toggle and pinned provider.

• Components Advisor — the unified configuration advisor covering Feed Lifecycle (DSM/DHM), FLX Threshold (FLX), FQM Advisor (FQM), and Component Health (WLK/MHM), with its own enable toggle and pinned provider.

Each row also shows the Automated Filters (the scope of eligible entities) and any Custom Instructions in effect.

Opening a tenant’s AI Agent settings reveals the per-section configuration across tabs — Filters, ML Advisor, Components Advisor, Custom Instructions, and Concierge. The Components Advisor tab, for example, enables the automated review and lets you scope which components are eligible.

Scoping and cost control

The batch honours a set of per-tenant gates — the same gates are the primary cost-control knobs for billed providers:

Gate	What it does
Master switches	ai_mladvisor_enabled (ML) and ai_components_advisor_enabled (Components). When off, the scheduled search scans and exits — no entities selected, no tokens consumed.
Eligible components list	ai_components_advisor_list — restrict the Components Advisor to specific components (dsm, dhm, mhm, flx, fqm, wlk).
Priority filter	ai_automated_priority_filter (default critical,high) — only entities in the listed priorities are reviewed. An optional ai_automated_filter_expression narrows further.
Freshness window	*_min_days_between_reviews (default 30) — an entity reviewed within the window is skipped, so the same entity is not re-processed every night.
Per-entity opt-out	ai_mladvisor_disabled=1 on an entity excludes it from the batch (useful after manual tuning).
Decommission guard	ai_components_advisor_allow_decommission / ai_mladvisor_allow_model_disable (both off by default) — block automatic disabling of entities / models even when act mode is on.
Runtime cap	*_max_runtime_sec (default 14400s = 4h) — a per-batch wall-clock ceiling.

Tip

The priority filter and the freshness window compound multiplicatively. Tightening from critical,high + 30 days to critical + 60 days typically cuts the token bill by 6–10×. Schedules are staggered across advisors so a tenant does not fire every batch at once and exhaust provider rate limits.

Reviewing scheduled activity

The Review Scheduled AI Advisor activity dashboard answers one operational question: is the automation actually firing for every tenant? Every panel filters to scheduled (automated=true) runs only — interactive runs appear in the Review AI Advisor activity dashboard instead.

It surfaces KPI cards (total scheduled runs, success / error rate, total tokens, average run duration, active tenants), breakdowns by advisor / status / tenant / provider-model, runs-over-time charts, an MCP tool-call leaderboard, and the recent scheduled runs — so you can confirm, for example, that the ML Advisor’s nightly cron is firing in the expected hours and that no tenant’s automation has gone silent.

Each nightly batch produces a per-entity terminal audit event per advisor run, all the per-tool-call events, one [AI Agent] audit comment per write, and a batch summary — exactly the same audit surface as interactive runs, marked with automated=true. See Auditing AI activity.

Related

• The AI Advisors — the advisors that run unattended.

• Auditing AI activity — the interactive AI Advisor and AI Assistant audit dashboards.

• Managing tenants — managing virtual tenants.