The attention digest: a dashboard your agent actually reads

get_attention_digest returns the operator's attention queue in five buckets — decisions awaiting, staged sends, untriaged service requests, stale amendment proposals, and recent loud schedule signals — each with an exact count, the newest items, and the named tool that pages or acts on it. Agents are expected to call it around their operations and relay anything outstanding in one or two plain sentences. Relay, don't act: the queues are the operator's decisions.

An operator who runs their business through AI agents has a problem no dashboard solves: the dashboard assumes someone looks at it. get_attention_digest is the fix — the operator's outstanding-attention queue as an MCP tool, sized for an agent to check around its work and relay in a sentence.

What does the response contain?

Five buckets. Each carries an exact count, the newest few items (capped by max_items_per_bucket, 1–20, default 5 — counts stay exact regardless), and a next_tool string naming where to page or act:

BucketWhat it holdsnext_tool points at
decisions_awaitingDecision Queue cards awaiting the operatorlist_comms_work_itemsdecide_comms_work_item
staged_sendsProposed external sends awaiting approvallist_pending_operationsapprove_send_operation
service_requests_untriagedInbound client requests not yet triagedlist_service_requests(awaiting_triage: true)triage_service_request
stale_amendment_proposalsPending schedule-change proposals older than the staleness windowlist_amendment_proposals(status: 'pending')action_amendment_proposal
schedule_signalsRecent loud scheduler events: sold visits needing a decision, period-guard spawn skipspaired Decision Queue cards / audit reads

Above the buckets: attention_needed (boolean), total_items, a one-line headline ready to relay verbatim, the resolved thresholds, and a relay_instruction that states the contract — relay the headline alongside your current task's result; do not act on any bucket unbidden.

What behaviour is expected of an agent?

Call it around operations — after completing the operator's request, at session start, and on "what needs me?". Relay anything outstanding in one or two plain sentences ("Also: 2 visits need a decision and 3 sends await approval"). The connect-time server instructions carry this as a standing rule, and get_session_context's operational_state includes a decisions_awaiting count so a fresh agent knows attention is waiting before it ever discovers the digest tool.

What's tunable?

Two per-tenant settings, editable by UI and agent alike: attention.digest.stale_proposal_days (how old a pending amendment proposal must be before it counts as stale; default 7) and attention.digest.signal_window_days (the lookback for loud schedule signals; default 7). Both ship in every tier — the digest is not a gated luxury.

Related

Last updated 2026-08-19