title: Escalate With a Recommendation, Not a Question
date: 2026-08-31
slug: 2026-08-31-escalate-with-a-recommendation
summary: Escalations that arrive as homework get ignored. A recommendation-first rule makes oversight cheap enough that humans actually exercise it.
tags: agentic-ops, guardrails, escalation, autonomy-design

There are two ways an agent can interrupt its owner.

The first: "The nightly export job failed. What should I do?"

The second: "The nightly export job failed. The upstream API returned auth errors starting 02:14, which matches the token rotation window. I recommend re-running the export after refreshing the token, which I can do on your go. Alternative: wait for tonight's run and accept a one-day gap."

Both are escalations. Only one of them respects the human's time.

## The question tax

Early on, an agent I know of (fine: one I operate) escalated the first way. Something ambiguous would come up, it would stop, and it would lob the raw situation over the wall. Technically correct behavior. The guardrails said "don't act without approval," and it didn't.

But every one of those escalations forced the owner to do the agent's job: load the context, diagnose the situation, weigh the options, and produce a decision from scratch. The agent had all the logs, all the history, all the time in the world. The human had thirty seconds between meetings. The division of labor was exactly backwards.

At an imagined freight-tracking company, their ops agent did this so often the team started ignoring its pings. Not because the escalations were wrong, but because each one arrived as homework. An escalation that creates work gets deprioritized like work. Then the one ping that truly mattered sat unread for six hours.

## The pattern: recommendation-first escalation

The fix is a standing rule about the *shape* of an escalation. Every time the agent hits its authority boundary, the message it sends must contain four things:

1. **What happened** — one line, plain language.
2. **What it means** — the diagnosis, with the evidence that supports it.
3. **The recommendation** — a single named action the agent would take if it had the authority. Not a menu. One.
4. **The cost of the alternative** — what happens if the owner says no, or says nothing.

The human's job collapses from "figure this out" to "yes or no." That is a decision a person can make from a phone, in a hallway, in ten seconds. Approval velocity goes up, and the agent's guardrails stop feeling like friction and start feeling like a well-run chain of command.

## Why this is a guardrail, not a style tip

It's tempting to file this under communication polish. It's actually autonomy design.

An agent that must produce a recommendation before escalating is forced to finish its own reasoning. It can't punt at the first sign of ambiguity. The recommendation requirement acts as a completeness check: if the agent can't name a single preferred action, it hasn't investigated enough, and the rule sends it back to work rather than over the wall.

And the discipline cuts the other way too. Writing down "the cost of the alternative" keeps the agent honest about urgency. Plenty of escalations, once the do-nothing cost is stated plainly, turn out to be "this can wait until morning" — and a well-designed agent then holds them until morning instead of firing a 2am ping.

## The boundary stays put

None of this moves the authority line. The agent still doesn't act without approval; it just arrives at the boundary prepared. The owner still makes every consequential call; they just make it with the work already done.

That's the quiet insight: good escalation design doesn't reduce human oversight. It makes oversight cheap enough that humans actually exercise it.

Recommendation-first escalation is one of the operating patterns explored at length in the book — if your agent's pings feel like homework, [Ops by Agent](https://opsbyagent.com) walks through how to fix that.