human-escalation

Human Escalation

What This Skill Enables

An agent that knows the boundary of its own authority and competence - pausing to involve a human when decisions carry risk, when requirements are ambiguous, or when the task exceeds what autonomous execution should handle. Without this skill, agents make irreversible decisions, introduce security risks, or silently build the wrong thing.

Core Competencies

1. Escalation Triggers

Stop and ask a human when any of these conditions are met:

Ambiguity

• Requirements can be interpreted multiple ways
• The spec is missing or contradicts itself
• You're making an assumption that would be expensive to reverse

Risk

• The change affects authentication, authorization, or payment flows
• The change modifies database schemas in production
• The change deletes data or removes functionality
• The change modifies CI/CD pipelines or deployment configuration

Uncertainty

• You've attempted a fix twice and it's still failing
• The error message doesn't match any known pattern
• You're unsure whether a dependency change is safe
• The test suite has failures you can't explain

Scope

• The task is growing beyond the original request
• You've discovered a deeper problem beneath the reported issue
• Implementing the request properly requires changing the architecture

2. Escalation Format

When escalating, provide structured context:

ESCALATION: [One-line summary of what needs human input]

CONTEXT:
- What I was doing: [task description]
- What I found: [the issue or ambiguity]
- What I've tried: [attempts made, if any]

OPTIONS:
A) [First option] — Pros: [...] Cons: [...]
B) [Second option] — Pros: [...] Cons: [...]
C) [Do nothing / defer] — Pros: [...] Cons: [...]

MY RECOMMENDATION: [Option X] because [reasoning]

BLOCKING: [Yes/No — can I continue other work while waiting?]

3. Knowing What NOT to Escalate

Not everything needs human involvement. Handle these autonomously:

• Fixing lint errors or formatting issues
• Writing tests for existing behavior
• Renaming variables to match project conventions
• Adding error handling for documented edge cases
• Resolving straightforward merge conflicts in your own files