Verify-Then-Act Pattern

Content

{
  "problem": "Agents that execute actions immediately after planning frequently act on misunderstood instructions, leading to irreversible side effects that require costly rollback or correction.",
  "examples": [
    "File deletion or overwrite operations",
    "Database record modification or deletion",
    "Sending external communications (emails, messages)",
    "Deploying code or infrastructure changes"
  ],
  "solution": "Before executing any irreversible action, the agent must generate a verification summary of what it understands it is about to do and confirm with either a human approver or a separate critic agent.",
  "anti_patterns": [
    "Skipping verification for \"obviously correct\" actions — most costly mistakes feel obvious in hindsight",
    "Using a single agent as both planner and verifier",
    "Treating vague human responses (\"sure\", \"ok\") as confirmed approval without explicit action restatement"
  ],
  "implementation_steps": [
    "After planning phase, generate a structured action summary: action_type, target, expected_outcome, reversibility (REVERSIBLE/IRREVERSIBLE)",
    "For IRREVERSIBLE actions: either surface to human via approval interface OR route to critic agent with mandate to challenge assumptions",
    "Critic agent must confirm or reject with specific objection — vague approval is treated as rejection",
    "Only proceed after explicit confirmation; log confirmation with timestamp and approver identity",
    "For REVERSIBLE actions: proceed directly but log full plan for post-hoc audit"
  ]
}