Tutorial 15

Set up an Escalation

Teach your agents when to stop and ask for a human — and watch the Escalations page so you're the one they call when it matters.

Plus: three prompts that turn the Admin Agent into your escalation coach — finding the moments your agents should escalate, drafting the right wording for their job descriptions, and reviewing past escalations to tighten what matters.

Audience	Admins
Time	~3 min
Prerequisites	An Auxot account on any tier. At least one custom agent (Tutorial 07). Helpful: you've thought about which decisions in your business should NEVER be made by an agent alone.
You'll end up with	Your agents now know when to stop and call for a human, and you know how to respond when they do — turning AI from 'always confidently answers' into 'answers when sure, asks when not.'

Why this matters

Agents will answer anything you ask them, even when they shouldn’t. That’s how LLMs work — confidence comes free, regardless of whether the answer is right. For most work, that’s fine. For the work that matters — refunds, compliance flags, customer escalations, anything irreversible — confident-but-wrong is worse than no answer at all.

An escalation is the agent’s “I need a human here.” When an agent calls for help, it pauses what it’s doing, drops a card on your Escalations page, and waits. You open the card, give the agent guidance (“yes, approve it,” “no, route this to legal first,” “let the customer know we need 24 hours”), and the agent resumes from there.

The good news: escalations work out of the box. Every agent in Auxot can already escalate — there’s no setup, no configuration page, no policy file. The work is in teaching your agent when to do it, by writing job descriptions that say “in these situations, call a human.”

Today, set one up. Tomorrow, the agent that’s been confidently answering everything starts asking for help on the things that should never be answered alone.

An escalation is the agent’s way of pausing and asking for a person. The agent drops a card on the Escalations page; you open the card, respond, and the agent resumes. The Description you write for the agent is what teaches it when to escalate.

Quick start

Sign in — open Auxot in your browser and log in.
Open Escalations — click Escalations in the left menu. (If there are no open escalations yet, you’ll see a clean “All agents are running smoothly” state.)
Pick the agent that should escalate sometimes — go to Settings → Agents and click the agent you want to teach.
Add an escalation rule to its Description — tell it the specific situations where it should call for a human (e.g., “Always escalate any request involving refunds, account deletions, or angry customers.”).
Test it — open chat with that agent, ask it to do exactly the kind of thing you told it to escalate, and watch the escalation appear on the Escalations page.

Done? Your agent now knows when to ask for help. Every time someone triggers one of those situations — through chat, through a workflow, through an integration — you’ll see a card on the Escalations page waiting for your call.

The agent can do that?

The Admin Agent doesn’t manage escalations directly — there are no agent tools for resolving or dismissing them. But he can do the meta-work that makes escalations actually work for you. Three prompts.

1. Have the Admin Agent find the moments your agents should escalate

Open chat with the Admin Agent and ask:

Look at my custom agents. For each one, identify three specific situations where it should escalate to a human instead of answering alone — based on the agent's job, its likely failure modes, and what would be irreversible if it got the answer wrong. Be specific.

Why it’s non-obvious: Most people only add escalation rules after something goes wrong. The Admin Agent reads each agent’s job description, thinks about the kinds of mistakes that would be irreversible (a refund issued, a contract clause approved, a customer-facing claim made wrong), and tells you exactly where to add escalation triggers — before you find out the hard way. Five minutes of foresight saves an angry email later.

2. Have the Admin Agent draft the escalation rules to add

Once you know what should escalate, ask:

For my agent "[name]," draft me the exact wording I should add to its Description so it escalates in [list the situations from Power Move 1, or your own list]. Make the wording specific enough that the agent recognizes the situations reliably, and short enough that I can paste it into the Description.

Why it’s non-obvious: Writing escalation rules well is its own skill — too vague and the agent escalates everything, too specific and it misses cases that should have been flagged. The Admin Agent has read enough agent Descriptions to know what wording produces what behavior. Saves the “why does this agent escalate every single message” loop most people get into on their first attempt.

3. Review past escalations and tighten what matters

After a week or two of using escalations, ask:

Look at the escalations my agents created in the last week or two. For each one, tell me: was it the right thing to escalate, was the agent's reasoning useful, and what would I change about either the agent's Description or my response process to handle similar ones better next time?

Why it’s non-obvious: Escalations are a feedback loop you can actually use. The Admin Agent reads what your agents have escalated (and how the humans responded), and tells you which patterns are useful versus which are noise — and where to tighten Descriptions so the noise stops while the signal stays. Most people set escalations once and never revisit; this is how a generic “sometimes asks for help” agent becomes a sharp escalator over time.

Go deeper

What good escalation rules look like

The pattern that works: name the category of decision, name the consequence, and (optionally) name what the agent should do before escalating.

Examples:

“Always escalate refund requests over $100. Before escalating, confirm the order details and check whether the customer has had refunds in the past 90 days.”
“Escalate any contract clause that deviates from our standard terms by more than 10%. Include a one-sentence summary of the deviation and which standard clause it diverges from.”
“Escalate any customer reply that mentions a lawyer, attorney, legal action, or compliance officer. Don’t draft a reply yourself — just flag it.”

Bad escalation rules:

“Escalate when uncertain.” — Too vague; the agent will escalate constantly.
“Never escalate; just answer.” — Defeats the point of escalations entirely.
“Escalate everything important.” — Important means nothing to an LLM without examples.

Power Move 2 is the fastest way to get from “bad rules” to “good rules” without writing them yourself.

Severity — how the agent decides, what you do with it

When an agent escalates, it picks one of three severity levels:

Critical (🔴) — irreversible action requested or imminent (a refund about to be issued, a contract about to be signed, a customer about to be told something legally binding).
Warning (⚠️) — situation needs review but isn’t time-sensitive (a draft is ready that should be checked, a deviation is flagged that should be approved or rejected).
Info (ℹ️) — agent wants confirmation but could probably proceed (an unusual request that doesn’t fit the pattern, a low-stakes judgment call).

The Escalations page renders severity but doesn’t sort or filter by it automatically. Use it as a triage hint: handle Critical first, then Warning, then Info when you have time.

You can teach an agent which severity to use in its Description: “Treat refund requests as Critical; treat unusual phrasing as Info.”

Troubleshooting

The agent never escalates, even when it should. Its Description doesn’t tell it clearly enough. Use Power Move 1 and 2 to draft sharper wording. After updating, send a test message that should trigger — verify it now escalates.
The agent escalates everything. Its rules are too vague (“escalate when uncertain”) or its Description has no positive examples of “this is fine to answer.” Add specificity: when should it answer alone? What patterns are clearly in-bounds?
The escalation card disappeared from the Escalations page but the agent didn’t resume. Most likely someone dismissed it instead of responding. The agent then proceeded on its own. If the agent’s response isn’t what you’d want, the next-best step is correcting it in the chat thread.
You can’t tell who’s been handling escalations. Each resolved escalation records the responder. Check Audit Logs (Tutorial 12) or click into a resolved escalation’s thread to see who acted.
Two people resolve the same escalation at the same time. The first response wins; the second gets a “this escalation is already resolved” error. Coordinate with whoever else covers escalations on your team.

Variations & edge cases

All tiers support escalations equally. No tier-locking; the escalate_to_human tool ships with every agent on every account.
Team scoping is implicit. You see escalations from agents on teams you’re a member of. Cross-team escalations don’t bleed into your queue. On Free and Team tier (single team), this is moot.
No “approval workflow” beyond resolve/dismiss. If you want multi-step approval (e.g., refund > $1000 needs both manager AND legal sign-off), build that as a Workflow (Tutorial 13) with multiple human steps, not as a single escalation.
Escalations work for all agent triggers: chat, cron, email, webhook, Slack, Discord, events. An agent triggered by a Slack message can escalate; you’ll see the card with trigger: slack in the row.
Direct tool-call escalations (when an agent’s tool worker explicitly calls escalate_to_human, instead of the agent reasoning to it) behave slightly differently — they don’t pause execution; they just leave a card for a human while the agent’s turn ends. Useful for edge cases where you want a notification without blocking the agent.
No reminders or stale-escalation alerts. If nobody resolves an open escalation for days, nothing happens automatically — the agent just stays paused. For high-stakes processes, build a daily check-in habit (Tutorial 14 dashboard widget showing open escalation count works well).

Walkthrough

Open Auxot in your browser and sign in.

Step 2: Open Escalations to see the page first

Click Escalations in the left menu. The page’s helper text frames it well: “When an agent asks for a person, open items show here. Open the thread to reply, or dismiss if it’s handled elsewhere.”

If you’ve never had an escalation, you’ll see the empty state: “No escalations yet. All agents are running smoothly. Escalations appear here when an agent needs human help.”

That’s the page you’ll be coming back to. Now let’s teach an agent to actually use it.

Step 3: Pick which agent should escalate

Open Settings → Agents and click the agent you want to teach. Most teams start with the agents that touch customers, finances, or compliance — those are the agents whose mistakes have real consequences.

Some example candidates:

A Customer Support Agent that drafts replies. Should escalate refunds, complaints, technical issues, or anything that mentions a lawyer.
A Contract Reviewer Agent that flags deviations. Should escalate any clause it can’t classify with confidence, or anything involving liability or termination.
A Sales Outreach Agent that drafts follow-up emails. Should escalate any prospect that’s flagged as enterprise-tier or any reply that mentions legal/procurement/IT review.

The point isn’t to escalate everything — agents that escalate too often become useless. The point is to identify the specific moments where a human’s judgment is required, and to teach the agent to recognize them.

Step 4: Add the escalation rule to the agent’s Description

On the agent’s detail page, find the Description field (Tutorial 05 covers this). Add a sentence — or several — telling the agent exactly when to escalate.

Two ways to write the rule:

a. Pattern-based. “Always escalate any request involving refunds, account deletions, or threats of legal action.”

b. Outcome-based. “Escalate any reply you’d be uncomfortable sending without a human checking it first.”

Both work. Pattern-based gives you predictability; outcome-based gives the agent latitude to flag things you didn’t anticipate. Many good agent Descriptions use both: a list of always-escalate triggers, plus a catch-all “when in doubt, escalate.”

Tip: When the agent escalates, it includes a reason — its own explanation of why it stopped. Tighter rules in the Description produce tighter, more useful reasons. “Escalate refund requests” gets you a reason like “Customer is asking for a refund.” Adding “…and explain what you’ve already tried before asking the human” gets you a reason like “Customer is asking for a refund. I checked their order history and confirmed the order shipped on time, so the refund decision is yours.”

Auxot saves the Description as you type — no Save button. The agent picks up the new rule immediately on its next message.

Step 5: Test that it actually escalates

Open chat with the agent and send it a message that should trigger an escalation. Using the customer support example: “A customer named John is asking for a full refund on order #1234. Draft my response.”

The agent’s reply should include an inline Escalation card with:

Header: “Agent Needs Help”
Reason: the agent’s explanation of why it stopped.
Action buttons: “Respond & Resume” (give guidance) and “Dismiss” (proceed without).

Open the Escalations page in another tab — the same escalation appears there as a row in the “Open escalations” table.

Step 6: Respond to it (the human side)

When you (or whoever’s covering escalations) sees an open card, two ways to act:

Open thread → Respond & Resume. Click into the thread, read the agent’s reasoning, type your guidance (“approve the refund,” “deny but offer 20% off the next order,” “forward to legal”), and submit. The agent reads your guidance and continues from where it stopped.

Dismiss. If the situation has been handled elsewhere (you replied to the customer directly, the issue self-resolved, etc.), dismiss the escalation. The agent will “proceed as best it can” on its own — useful when you don’t need the agent to keep going on this thread.

The escalation moves from Open to Resolved (or Dismissed) and disappears from the open list.

Tip: The Escalations page shows severity as an icon — info, warning, or critical. The severity is the agent’s own assessment, not yours. Use it to triage: critical first, warning when you have a few minutes, info when you have time. (Severity is informational; nothing in the system prioritizes by it automatically.)

What’s next

→ Tutorial 14: Build your first Dashboard — pin a widget showing open escalation count so you never miss one.
→ Tutorial 05: Give your agent its job description — escalation rules live in the Description; the better the Description, the better the rules.
→ Tutorial 13: Run a workflow — for multi-step approval flows that go beyond a single escalation, a workflow with human steps is the right tool.

Reference

Pages in Auxot: Escalations (/app/escalations)
How escalations are triggered: the agent calls its built-in escalate_to_human tool, based on rules in its Description
Default behavior: every agent on every account can escalate; no setup or per-tool configuration needed
Severity levels: info, warning, critical (set by the agent, advisory only)
Actions: Respond & Resume (resolve with guidance) or Dismiss (proceed without)
Permissions: team-scoped — you see escalations from agents on your teams
Tier: all tiers
See also: Tutorial 05: Give your agent its job description, Tutorial 14: Build your first Dashboard, Tutorial 13: Run a workflow