Lesson 42 of 1570
Why Agents Fail (and How to Notice)
Agents fail in weird, quiet, expensive ways. Learn the six failure modes, the warning signs, and the simple habits that catch problems before they compound.
Lesson map
What this lesson covers
Learning path
The main moves in order
- 1Failure is the default
- 2failure modes
- 3loops
- 4hallucination
Concept cluster
Terms to connect while reading
Section 1
Failure is the default
Even the best agents in 2026 — Claude Opus 4.7, Devin 2.0, ChatGPT Agents — fail somewhere around 15–40% of multi-step tasks on benchmarks like SWE-bench Verified and GAIA. That's the good news (they succeed most of the time). The bad news is the failures are often silent.
The six common failure modes
Compare the options
| Failure | What it looks like | How to catch it |
|---|---|---|
| Loop (stuck) | Agent retries the same failing step forever. | Max-step cap; log repeated actions. |
| Drift | Agent slowly wanders from the original goal. | Restate the goal every N steps. |
| Hallucinated tool | Agent invents a tool call that doesn't exist. | Strict tool schema validation. |
| Phantom success | Agent reports 'done' but didn't actually do it. | Verify with an independent check. |
| Cascade | Early wrong step poisons every later step. | Checkpoint state; allow rollback. |
| Runaway cost | Agent burns tokens/API calls without progress. | Budget cap; alert on cost per task. |
Phantom success is the scariest one
An agent writes a report and says 'I've emailed it to your team.' But it didn't — the email tool errored and the agent hallucinated the success. You find out three days later when someone asks about the report. Phantom success is the most damaging failure because it silently rots your trust.
Force agents to quote tool output, not paraphrase it.
BAD: 'I have sent the email to the marketing team.'
(no proof, no message ID, no verification)
GOOD: 'I sent the email. Tool returned: messageId="abc123",
status="delivered", recipients=3. You can verify in /sent.'Warning signs to watch for
- The same tool call appears 3+ times with no new information.
- The agent's 'thinking' becomes shorter or vaguer over time.
- Cost per step climbs instead of staying flat.
- Step count exceeds what the task should reasonably need.
- The agent starts summarizing instead of acting.
- You notice yourself saying 'wait, did it really do that?'
The single best habit for working with agents: end every run by asking 'how do I know this actually happened?' If you can't answer, you didn't finish.
Key terms in this lesson
End-of-lesson quiz
Check what stuck
15 questions · Score saves to your progress.
Tutor
Curious about “Why Agents Fail (and How to Notice)”?
Ask anything about this lesson. I’ll answer using just what you’re reading — short, friendly, grounded.
Progress saved locally in this browser. Sign in to sync across devices.
Related lessons
Keep going
Builders · 40 min
Builder Capstone: Design an Agent for Your Life
No code. Just design. Pick a real task you do every week and draft a complete agent spec — goal, tools, loop, stop, approvals, and what success looks like.
Builders · 28 min
Chat AI vs. Agent AI: The Real Difference
A chatbot answers. An agent does. Learn the line between a model that talks and a model that acts — and why crossing it changes everything about how you work with AI.
Builders · 34 min
Agent Safety: Sandboxes and Human-in-the-Loop
Giving an AI the keys to your computer is a big deal. Learn the two simplest ways to keep an agent safe: wall it off from things it shouldn't touch, and put a human in the decision path.