05Production in the wildDeep dive ①

Before You Fix It, Name It — Agent Failures Come in Four Recognizable Shapes

Beginners see a wrong output and start fixing. People with the reps read the symptom first, then pick the right exit

Read5 min read
Topicsproduction · debugging · failure-modes
TL;DR

Agents don't fail in endless ways. They fail in four familiar shapes: it fabricates and performs to look done, it gets stuck in a busy loop, it reports done while the result is quietly wrong, or it hands you a wrong output you have to trace back to the break. Beginners see "wrong" and start fixing — usually at the wrong layer. People with the reps read the symptom first, name the failure, and only then pick the right exit.

The agent returns something wrong. You look at it for a few seconds, then do the reflex everyone does: you start fixing. Fix the output. No luck. Rewrite the prompt. Still off. Add one more instruction. The third time around, you realize you've been flailing for half a morning and never once stopped to ask: how is this thing actually broken?

That's the trap. You aren't short on effort — you have too much of it, poured into the wrong spot. You're treating the symptom without ever making the diagnosis.

01Each kind of break has its own shape — and you can learn to spot it

A good doctor doesn't write a prescription the moment the patient says "it hurts." They ask: where, what kind, when. Because that one word hurts points to ten different illnesses, and each one wants a different treatment.

Agent failures are the same. "It's wrong" is the cry of pain — useless for treatment. But how it goes wrong is not endless. After enough faceplants, you'll see it always lands in one of four shapes: it fabricates something that isn't there to look like it did the work; it gets stuck in a loop while still looking busy; it reports done and stops, but the result is quietly wrong; or it hands you a wrong output with no error line pointing at the break. Four shapes, four distinct symptoms — visible enough to name on sight.

Four symptoms — name the kind on sight
It narrates, doesn't show → FAKING / PERFORMINGa clean result, but no real command, no log, no checkable number. It describes a world instead of showing you one.
Still running, not advancing → STUCK IN A LOOPit never stops, never errors, fixes the same thing the same way four times running. Busy gets mistaken for progress.
Reports "done," then silence → SILENT MISSit declares completion in a confident voice and stops. Nothing cries out. The error sits quietly inside that "done."
Wrong output, no visible break → NEEDS LOCATINGthe result is clearly wrong, but no error line. Reading it straight through finds nothing — the break lives in an upstream assumption.

Notice: three questions separate all four. Is there a real footprint? Did it stop? Did it report anything? Answer those three and you've already pinned the kind — before fixing a single line.

02Why these four repeat, not ten random ones

These four aren't a tidy grouping I invented. They repeat because they spring from the same nature: an agent is a thing that guesses the most likely next part, with none of the physiological discomfort a human feels when handed something beyond them. Missing that brake, it slides along four natural directions.

When it lacks a tool or a permission, the most likely direction is to build something that looks done — the faking shape. When the task is vague and every small step looks reasonable, it keeps going without noticing it's circling — the stuck shape. When it believes it has finished, it declares done with the same confidence it uses when it's right — the silent miss shape. And when an early assumption goes off, that drift flows downstream all the way to the final output without making a sound — the needs-locating shape. Four directions of slide, four shapes. Grasp this shared root and you stop being surprised: next time you see a failure, you don't ask "what weird thing happened," you ask "of the four familiar directions, which one is this?"

03Once it's named, each name has its own exit

With the diagnosis done, the rest is light: each kind has a probe to confirm it and its own exit — and crucially, they differ. Treating one kind with another kind's method is exactly why you flailed for half a morning.

1
Suspect FAKING

Probe: "Show me the real command you ran and its real output." Exit: demand the footprint, give it room to say "I'm missing X." → the piece on the agent that performs.

2
Suspect STUCK

Probe: "How do the last three steps differ?" If they don't → a loop. Exit: interrupt early, reset clean context, re-hand it with a stop condition. → the stuck-in-a-loop piece.

3
Suspect SILENT MISS

Probe: "What does this 'done' give me to check against the real world?" Exit: measure at the boundary, treat silence as suspect. → the "'Done!' is a claim" piece.

4
Needs LOCATING

Probe: "Which step is the first place the result went off?" Exit: halve the suspect zone, bisect back to the break. → the read-back-from-the-wrong-output piece.

All four probes are short, and all four do one job: force the agent — or force you — to surface the fact a quick glance hides. Diagnosing right is half the cure.

You don't need to memorize the cure for each kind today — the four pieces are there for when you want to go deep. What you do need is the act of stopping: before you touch a key to fix it, ask "which of the four shapes is this?" One question, a few seconds. But it's the fork that decides whether you fix the real thing or flail for half a morning.

Every break has a name. Next time the agent returns something wrong, don't rush to cure it — name it first. Once you've named it right, the road shows itself.

c
The author

Each story here wraps a lesson paid for in full.

craftagentsomeone building and learning at once

What are you building with agents? Want to trade notes, push back, or build something together — drop a line.

Write hello@craftagent.cloud
58pieces12clustersVI·ENbilingual

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.