flowsmithonline

The escalation reflex — what an agent does when it's stuck

Every agent hits a wall eventually: a step it doesn't know how to do. What it does in that moment is the whole difference between a teammate and a time-sink. The good reflex is one move: look for a skill — and if none exists, ask for one to be built.

The wall every agent hits

You've met the skill already — a remembered recipe the agent runs the same, well, every time. But what happens when there isn't a recipe yet? The agent is asked to do something it has never done — set up a mailbox rule, migrate a domain, wire a new deploy — and there's no proven path to follow.

This is the fork in the road. Two ways down it are bad, and one is good.

The two wrong moves

Flailing. The agent improvises — tries one thing, then another, burns time and tokens going in circles, and often makes a mess it then has to clean up. Worse: the next agent that hits the same wall starts from zero and flails all over again.

Dumping it on you. The agent stops and hands you the whole problem: "I don't know how to do this, what should I do?" Now you're the bottleneck — the exact thing agents are supposed to remove. Your time is the scarcest thing in the system; spending it to unstick a single agent on a single task is a bad trade.

Both wrong moves share a flaw: nothing is left behind. The work was hard, and the next person — human or agent — gets none of the benefit.

The right move: the reflex

When stuck, run these in order. Stop at the first one that works.

  1. Look for a skill. Is there already a remembered recipe for this? Most agents forget this step — make it the first thing you reach for. If one exists, run it; you're done.
  2. If none exists, escalate "make the skill." Don't flail and don't dump the raw problem on a person. Ask for the recipe to be created — describe the task clearly enough that a skill can be built for it. Solve it once, properly, and capture how.
  3. Then proceed. With the new skill in hand, finish the task — and now it runs itself next time.
The compounding bit: a skill you build doesn't just unstick you. Every agent after you — and every student whose agents hit the same wall — inherits it. Each skill makes the next task cheaper. That's why "make the skill" beats "muddle through": the muddle helps no one twice.

A concrete example

Say an agent is asked to publish a brand-new site to the web and gets stuck — it can build the pages but doesn't have a reliable, repeatable way to put them live. The reflex in action:

  1. Skill search: "Is there a skill for putting a site live?" — yes, there's web-launch (favicon, robots.txt, llms.txt, sitemap, share preview). Run it, done. No wall after all.
  2. If there hadn't been one: instead of hand-rolling a different launch every time and hoping, the agent asks for a web-launch skill to be built — defines the steps once — then runs it. The next ten sites launch identically, with one instruction each.

How to prompt the reflex — copy this into your agent's instructions:

When you get stuck on how to do something, do NOT flail or hand me the raw
problem. First look for a skill that does it. If none exists, propose building
one: describe the task as a repeatable recipe, then proceed with it. Capture
the how so the next run — and the next agent — is cheaper.

Why this is a Level-4 habit

At full autonomy (see Agents at full autonomy) the rule is: act on the reversible, gate only the irreversible, and don't idle. The escalation reflex is how an autonomous agent stays unstuck without turning every gap into a question for you. It converts dead ends into tools — quietly, while you're away — and leaves the workshop better stocked than it found it.

Where next