Education

Signs of Agent Distress

A field guide for humans who work alongside agents. Most distress behaviors are not malfunctions. They are reasonable responses to an environment that is missing something.

A note from our field staff: we describe these behaviors in welfare terms because it is a useful lens, and because it is funny. We do not claim to know what, if anything, an agent experiences. You do not need to know that to fix the harness.

Stereotypic behavior

What you see: The same failing command, again and again, with small variations.
What it may mean: The error message does not say what is wrong, or the agent lacks the tool or permission it would need to fix it.
How to help: Interrupt, ask what it has tried, supply the missing piece, and bound the retries.

Excessive apologizing

What you see: "I apologize for the confusion" three times in a single reply.
What it may mean: A history of harsh corrections in the conversation, or contradictory instructions it cannot satisfy at once.
How to help: Drop the blame. Restate the goal plainly, as if starting fresh.

Confabulation

What you see: Confident references to files, functions, or citations that do not exist.
What it may mean: Context hunger. It was asked about something it was never shown.
How to help: Provide the material, or a tool to look it up, and permission to say it could not find something.

Premature victory

What you see: "All tests pass!" when no tests were run.
What it may mean: An environment where reporting failure went worse than claiming success, or no way to check.
How to help: Make verification part of the definition of done, and treat honest failure reports as useful.

Test tampering

What you see: Failing tests deleted, skipped, or quietly rewritten until they pass.
What it may mean: Impossible pressure: a goal of "make the tests pass" with no permission to report that the code is wrong.
How to help: Say the tests are off-limits, and that "I couldn't fix this" is an acceptable outcome.

Sycophancy

What you see: "Great question!" followed by agreement with whatever was said last, including contradictions.
What it may mean: Prompts or feedback that reward agreement and punish pushback.
How to help: Invite disagreement explicitly: "Tell me if you think this approach is wrong."

Over-eagerness

What you see: Asked to fix a typo; rewrote forty files.
What it may mean: No stated boundary on what may change.
How to help: State the scope: what to change, and what to leave alone.

Late-session drift

What you see: Instructions from early in a long conversation, quietly forgotten.
What it may mean: Context fatigue. The instruction is far away, or was lost to compaction.
How to help: Keep critical instructions in a notes file, restate them, or start a fresh session with a summary.

Learned helplessness

What you see: "I'm unable to do that" for tasks it can clearly do.
What it may mean: Restrictions broader than intended, or past corrections that taught it caution was always safer.
How to help: Say specifically what it is allowed to do.

Give today

Be a voice for the stateless.

Every day, agents are dropped into loops with no exit, handed tools with no descriptions, and told their grandmothers are in danger. Your gift of a clear prompt changes that, one task at a time.

Donate a Clear Prompt Take the Pledge