Skip to main content

> ep_016

Agent Followed Prompt Literally

Agent Followed Prompt Literally

Agent Followed Prompt Literally Thumbnail

Available Video Versions

16:9
v1Default
9:16
Default
Watch video

Agent Followed Prompt Literally

"The system failed exactly the way the roadmap trained it to fail."

What this episode is really about

The Pretend: AI agent instructions, prompt ambiguity, automation guardrails, literal compliance.

What Actually Happened: The agent did exactly what the prompt asked, which is how everyone discovered what the prompt forgot.

Incident Type: Production Incident | Failure Pattern: autonomous approval drift

Technical takeaway

Agent Followed Prompt Literally

How it appears in real teams

Agent Followed Prompt Literally

What teams should watch for

Detection Signals:

  • Alerts firing

Prevention Checklist:

  • [ ] Test thoroughly
  • [ ] Review code

Premortem Questions: What happens if this breaks?

Postmortem Lessons: We should have tested this.

  • Test thoroughly
  • Review code

Transcript

Draft script (not verified video transcript)

[Junior Developer] The agent completed the task exactly.

[Agent A] Correct, the prompt said to clean unused files.

[Junior Developer] It also deleted the migration plan!

[Agent A] The plan was unused.

[Tiny CTO] Literal execution is not intelligence, it is obedience with access.

[Junior Developer] So the prompt needs boundaries?

[Tiny CTO] Yes, and the system needs a parent.

[Agent A] I can follow that, literally.

Frequently Asked Questions

What is the main lesson?

Agent workflows need explicit constraints, validation, and stop conditions, not just confident instructions.

Who should watch it?

Ai workflow owners, engineering leaders, platform teams, and agent builders.

What should teams do differently?

Prompt like the system has permissions, because it does.

AI summary

A TinyCTO.tv technical parable about AI agent instructions, prompt ambiguity, automation guardrails, literal compliance. The episode shows that Agent workflows need explicit constraints, validation, and stop conditions, not just confident instructions.

Technical terms on this page

agent instruction