How Agents Fail

Understanding failure modes is the first step to fixing them.

The Five Failure Categories

1. Comprehension Failures

The agent misunderstands what you want.

Symptoms:

• Does the wrong thing confidently
• Solves a different problem
• Misses key requirements

Example:

You: "Delete the old logs" Agent: deletes all logs including today's (You meant logs older than 7 days)

2. Memory Failures

The agent forgets or misremembers.

Symptoms:

• Forgets previous context
• Contradicts earlier statements
• Loses track of task progress

Example:

Agent: "I'll save this to NOTES.md" Later Agent: "I don't have any record of that"

3. Tool Failures

Tool calls fail or return unexpected results.

Symptoms:

• Commands error out
• Wrong tool used
• Tool output misinterpreted

Example:

Agent calls: exec("npm install") Error: EACCES permission denied Agent: "Installation complete!" (didn't read error)

4. Reasoning Failures

The agent's logic is flawed.

Symptoms:

• Wrong conclusions from correct data
• Circular reasoning
• Missing obvious solutions

Example:

Agent: "The build failed because tests failed" Agent: "Tests failed because the build failed" (Circular, unhelpful)

5. Hallucination Failures

The agent makes things up.

Symptoms:

• Cites non-existent sources
• Invents data/statistics
• Claims to have done things it didn't

Example:

Agent: "According to the 2024 Smith Report..." (No such report exists)

The Common Thread

Most failures come from:

1. Unclear instructions — Agent doesn't know what you want
2. Missing context — Agent doesn't have needed information
3. Poor system design — Architecture doesn't support the task

If your agent is wrong, your system design is probably wrong — not the AI.