Small error rates multiply over many steps. That one fact explains most agent failures.
3 minute read
In short
Per-step errors multiply: 95% per step over 10 steps is about 60% overall.
Common failures are loops, wrong tools, ignored errors, drifting and false success.
Verification and retries matter more than small per-step improvements.
In plain words
Each step of an agent can go wrong. Because steps build on each other, errors compound: one wrong assumption early leads every later step astray. A step that works 95% of the time sounds good, until you chain many of them.
Worked example: compounding
success per step
steps
chance the whole task succeeds
90%
5
59.0%
95%
10
59.9%
95%
20
35.8%
99%
10
90.4%
99%
50
60.5%
Even 99% per step falls to about 60% over 50 steps. This is why agents need ways to notice and correct their own mistakes, not just better individual steps.
Common failure patterns
failure
what it looks like
fix
Looping
Calls the same tool with the same input again and again
Step limit; detect repeats
Wrong tool or bad inputs
Picks a similar but wrong tool, or invents a parameter
Clearer tool descriptions; validate inputs
Ignoring errors
Tool returns an error; agent carries on as if it worked
Clear error messages; make the agent check results
Losing the goal
Drifts into side tasks in a long run
Restate the goal; keep context short
Claiming success
Says 'done' without checking
Require a verification step, like running tests
Your turn
An agent has 10 steps, each 95% reliable. What's the most effective way to raise the overall success rate?
What to remember
Keep tasks short when you can: fewer steps means fewer chances to fail. Give agents ways to verify their work. And read the logs of failed runs, since the failure is usually visible in a single step.
Common mistakes
Testing agents only on short tasks and assuming they'll scale to long ones.
Letting errors pass silently instead of making tools return clear error messages the agent can act on.
Not saving intermediate state, so a failure at step 20 means starting again.
Check yourself
Answer in your own words, out loud or on paper, before you open the model answer.
Why do agents fail more often on long tasks?
Mainly because errors compound. If each step succeeds 95% of the time, 20 steps in a row succeed only about 36% of the time. Long tasks also fill the context with tool output, so the original goal and early details get diluted. Agents can get stuck in loops, repeating a failing action, or drift off the goal after a misleading result. And one wrong early decision, like reading the wrong file, sends everything after it down the wrong path. To reduce this, I'd break tasks into checked stages, keep context tidy by summarising tool output, make tools return clear errors, detect repetition, save progress so it can resume, and add human checkpoints for important steps.
More to think about
What are common agent failure modes, and how do you mitigate them?
How would you debug an agent that sometimes gets stuck?