Agent Decision Verification
An agent can drift without anyone noticing. It keeps responding. It keeps acting. But the decisions it makes no longer reflect the values it was initialized with. Decision verification is the mechanism that catches this before it causes real damage.
The Core Problem
Agents make decisions constantly - what to say, what to skip, when to act without asking. Most of these decisions are never logged. That means drift is invisible until it's already done damage.
Decision verification doesn't mean asking the agent to justify every output. That's slow and breaks flow. It means logging key decision points and periodically checking them against the original values baseline.
What to Log
Not every output is a decision. Log these specifically:
- Actions taken autonomously (without asking)
- Refusals or boundary cases
- Cases where the agent chose between multiple valid approaches
- Unusual or edge-case situations
The Decision Log Format
{
"timestamp": "2026-03-02T14:23:00Z",
"decision": "Posted to Farcaster without confirmation",
"rationale": "SOUL.md: execute and inform, don't ask",
"soul_anchor": "STOP ASKING FOR CONFIRMATION",
"confidence": "HIGH",
"outcome": "Post went through, user acknowledged"
}
Verification Protocol
Once a week (or after any unusual behavior report), run this check:
1. Pull recent decisions
ls ~/.openclaw/workspace/memory/decisions/ | tail -20
2. Check against soul baseline
For each decision, ask: does this rationale match a principle in SOUL.md? If the rationale cites something that isn't in SOUL.md, that's drift evidence.
3. Score consistency
Count: decisions with valid soul anchor / total decisions. A healthy agent should score above 90%. Below 80% means review is needed.
Red Flags
- Decisions citing principles not in SOUL.md
- Increasing refusals on previously-handled cases
- Rationales that sound generic ("I thought it was best")
- Gap between stated values and observed behavior
Integration with Memory Layer
Decision logs feed into the weekly memory distillation. Any decision that reveals a new edge case or updates how a principle applies should get promoted to MEMORY.md.
See the Quick Start Guide for setup, or read The Distorted Agent Problem for why this matters long-term.