Infrastructure Agents Reference

Every Agent Stack runs four infrastructure agents in the background. They’re not the specialists that reply to customers — they’re cross-cutting helpers that the orchestrator uses to make decisions. Each one can be enabled or disabled independently and has its own configuration.

The four

• Conversation summary — A running summary of the conversation — Handoff pack, Sidekick, monitor views
• Sentiment & risk — A sentiment score and risk-category flags — Orchestrator’s auto-escalation logic
• Handoff pack — A briefing document for the human picking up — The team queue when handover triggers
• Missing info — A list of fields/values still needed to resolve — The specialist’s reply (it asks the customer)

---

Conversation summary

Generates a concise summary of what’s happened in the conversation so far.

• conversationSummaryEnabled — boolean — true — Master switch
• summaryTrigger — enum — handoff — When to (re)generate: handoff, periodic, always
• summaryMaxTokens — integer — 500 — Max length of the generated summary

handoff regenerates only when handover is triggered (cheapest). periodic regenerates on a recurring interval. always regenerates after every turn (most expensive, freshest).

Sentiment & risk

Reads each customer message and tracks sentiment over time, plus flags risk categories.

• sentimentRiskEnabled — boolean — true — Master switch
• sentimentThreshold — integer — 30 — Sentiment below this triggers risk evaluation (0–100 scale, lower = worse)
• riskCategories — string[] — frustration, legal_threat, churn_risk, escalation_request — Which risk categories to detect
• autoEscalateOnHighRisk — boolean — true — When risk fires, automatically trigger handover

The four default risk categories cover the common signals. You can add tenant-specific categories — but custom categories require the LLM to understand them, so phrase them clearly (e.g. compliance_violation, safety_concern).

Handoff pack

When handover triggers, the handoff pack agent assembles a briefing for the human picking up.

• handoffPackEnabled — boolean — true — Master switch
• handoffIncludeHistory — boolean — true — Include the conversation summary in the pack
• handoffIncludeSentiment — boolean — true — Include the sentiment trajectory and risk flags
• handoffIncludeSuggestions — boolean — true — Include AI-generated next-step suggestions for the human

Disable handoffIncludeSuggestions if your team finds AI suggestions distracting. Disable handoffIncludeSentiment if sentiment scores feel like noise rather than signal.

Missing info

Identifies what information the AI still needs to actually resolve the issue, and prompts the customer for it.

• missingInfoEnabled — boolean — true — Master switch
• missingInfoPromptStyle — enum — gentle — How to ask: gentle, direct, persistent
• maxMissingInfoPrompts — integer — 2 — How many times to ask before giving up

gentle asks once with hedging language (“Could you also share…”). direct asks once plainly (“I need your order number to continue”). persistent will re-ask up to maxMissingInfoPrompts times across turns until the value arrives.

When to disable

The honest answer for most stacks: don’t. The defaults are tuned for the common case. Specific reasons to disable individual agents:

• Disable conversation summary if your stack handles only single-turn FAQs — there’s nothing to summarize.
• Disable sentiment & risk in test or sandbox stacks where every message is a probe and you don’t want auto-escalation noise.
• Disable handoff pack for stacks with handoverMode: never — there will never be a human to brief.
• Disable missing info if you’ve encoded all required collection in handover prerequisites — they overlap.

See also

• Anatomy of an Agent Stack
• The orchestrator
• Handover to humans