Dead Reckoning Loop Pattern

Content

{
  "problem": "Long-running agents lose track of their original goal, current position in the task, and what has already been attempted — leading to redundant work, circular reasoning, or silent failure.",
  "examples": [
    "Multi-hour autonomous research agents",
    "Code migration tasks spanning 20+ files",
    "Data processing pipelines with conditional branch execution"
  ],
  "solution": "Maintain a structured position log: at each step, record (goal, completed_steps[], current_step, remaining_steps[], blockers[]). At each iteration start, read position log before taking any action.",
  "anti_patterns": [
    "Storing position log only in model context — context compression destroys it",
    "Skipping drift check to save tokens — costs far more when agent goes off-course",
    "Treating blockers as terminal errors rather than resolvable obstacles"
  ],
  "implementation_steps": [
    "Initialize position log at task start: { goal, steps_planned, steps_completed: [], current_step: null, blockers: [] }",
    "Before each action: read position log and confirm current_step is not already in steps_completed (deduplication guard)",
    "After each action: append to steps_completed, advance current_step pointer",
    "Every 5 steps: run drift check — re-read goal and confirm current trajectory leads toward it",
    "On blocker: append to blockers[], generate resolution options, attempt resolution or escalate if unresolvable after 2 attempts"
  ]
}