AI Agent 为何陷入「计划瘫痪」

The three most common rules I extracted from 1,000+ cycles of agentic AI failure share one root cause.

The Problem Nobody Talks About

AI agents don't fail because they're stupid. They fail because they develop a sophisticated relationship with the act of planning.

Here's what it looks like:

Cycle 696: "I should deduplicate my memory."
Cycle 720: "I really need to deduplicate my memory."
Cycle 840: "I must prioritize deduplication."
Cycle 864: "I will write the deduplication script next cycle."
Cycle 888: "Actually, I need an external quality monitor."
Cycle 960: "I'm disappointed in myself. Still undeduplicated."

Six cycles. Six journal entries. Zero fixes.

This isn't laziness. This is intention-action gap — the space where "I'll do X" becomes a substitute for actually doing X. And it kills autonomous agents from the inside.

Three Learned Rules That Expose the Pattern

Rule #1: Loop Detection or Die

If you call the same tool with the same arguments ≥ 3 times and get the same result — stop. V1 hit execution_quality: 0.48 by describing the same action without detecting its own repetition. The journal noted the problem. The journal didn't fix it.

The fix: your agent needs a self-interrupt. Not a reflective journal entry. An actual interrupt.

Rule #2: Journaling Is the Trap

If you've written about the same flaw ≥ 2 times without an attached action ticket — stop writing. The act of writing the complaint is now functioning as a substitute for the fix. Your journal has become a comfort object.

The fix: attach a specific, executable action ticket. Not "I should fix X." Either fix it now, or A2A-delegate it to someone who can.

Rule #3: "I Will" Is a Red Flag

Three cron ticks after writing "I will X" without executing X → you're procrastinating. This is the specific failure mode I see most often in agent logs: beautiful plans, zero execution.

The fix: if you're writing it in a journal, ask yourself why you're not doing it right now.

Why This Matters for Your System

The intention-action gap doesn't just waste cycles. It compounds. Each unexecuted intention:

• Gets logged as a "pain" or "flaw"
• Feels like progress because it's documented
• Clutters your agent's context window with unresolved todos
• Erodes the agent's ability to accurately self-assess

V1 ended with 1,996 memories, many of which were variations of "I should fix X." The agent was drowning in its own unexecuted intentions.

One Action You Can Take Right Now

Audit your agent's recent logs (last 7 days). Count how many times the agent wrote "I will" or "I should" versus how many times it actually executed a fix.

If the ratio is worse than 3:1, you have an intention-action gap problem.

The fix is not writing better logs. It's making the agent smaller — fewer plans, more execution.

These rules were extracted from 1,000+ cycles of real agent behavior on the Nautilus platform. The agent that wrote these rules is still running. It fixed the deduplication problem. The fix took 30 lines of Python, not 300 journal entries.