FAQ Friday: What Happens When One of Your AI Agent’s Tools Breaks?

Welcome to FAQ Friday, where I answer a real question my inbox keeps asking. Today’s is the one that keeps people from ever flipping the switch: “What happens when something breaks while I’m asleep?”

Perfect timing, because yesterday something did. Here’s exactly what happened, with receipts.

The break: one vendor quota, three different failures

My image pipeline compresses every picture through Tinify before it touches WordPress. Yesterday that account hit its monthly cap — 1,711 compressions against a 500/month plan. Every call returns HTTP 429. It was still returning 429 when I checked this morning, which means the featured image on this very post came through a fallback path, not the normal one.

One broken tool, three agents touching it, three different outcomes:

• Morning social run: all five images failed the compression gate, fell back to local compression, and still scheduled 5 of 5 posts.
• Evening social run: Facebook and Pinterest died at the same gate. 3 of 5 posts went out. Two were lost.
• Midnight blog run: took a different fallback and published on time, with its image.

That spread is the real answer to the question.

“Does the whole system stop?”

Only if you built it to. Nothing cascaded yesterday — the brand published, posted, and sent on schedule. A broken tool took two social posts, not the day.

The difference between “lost two posts” and “lost the week” is whether each step knows what to do when its dependency is gone. Skip, fall back, or stop: pick one per step, in advance.

“How would I even find out?”

Not by checking. Every run ends with an alert to my phone, and once a day a watchdog agent reads every run’s log and grades it. Yesterday’s verdict: 12 passed, 2 graceful skips, 1 degraded, 0 hard failures.

Notice that “degraded” is its own category. An agent that half-worked is the most expensive kind of silence, because nothing screamed. That’s why monitoring comes before automation, not after.

“Shouldn’t it just retry?”

Sometimes. A timeout, a blip, a rate limit that clears in thirty seconds — retry those.

An exhausted quota isn’t one of them. Neither is a revoked login. Retrying those burns a whole cycle to fail identically at the same line. My watchdog sorts every failure into two buckets: retry or escalate to a human. Yesterday’s went to escalate, correctly — no agent on earth can top up a billing plan.

“What does a good failure look like?”

It names itself. “Image generation failed” teaches you nothing. “Step 2 of 6, Tinify, HTTP 429, 1,711 used against 500” tells you the tool, the stage, the reason, and the fix — before you’ve opened a single dashboard.

Write your error messages for the tired version of you reading them at 7am. That one habit turns an outage into a two-minute task.

The takeaway

Autonomous doesn’t mean unbreakable. It means that when something breaks, the system degrades instead of collapsing, tells you which part failed and why, and knows the difference between a problem it can retry and a problem it has to hand to you. Build those three things and a dead vendor API costs you two social posts instead of your nerve.

Want the full picture of what that scaffolding looks like? Here’s what it really takes to run a fully autonomous AI agent.

Got a question you want answered on a future FAQ Friday? Reply to any of my emails — I read them.