scobleclaw/coach-pattern-analysis.md

Coach Critique Analysis: Recurring 5/10 "Repetitive Behavior Pattern"

Issue: marcus/scobleclaw #9 Date: 2026-05-23 Analyzed by: agent-cos

Summary of Findings

Three consecutive coach critiques (coach #24, #25, #26) all scored 5/10 overall with the same root complaint: agents looping/repeating instead of completing work. After thorough analysis, these critiques are valid signals, not noise. Three distinct agents (trust, engineering, vault) all exhibit the same failure mode.

---

Critiques Analyzed

#24 — Trust Agent (May 7)

• Scores: Accuracy 7, Efficiency 4, Tools 6, Skills 5, Overall 5
• Pattern: Repeated identical "trust skip" entries at 14:23 and 15:23 PT with no new analysis; yields on all three ventures repeatedly
• Root cause: When all ventures are blocked on human, agent enters yield loop with no escalation mechanism

#25 — Engineering Agent (May 10)

• Scores: Accuracy 7, Efficiency 2, Tools 6, Skills 4, Overall 5
• Pattern: Repeatedly states inability to write files, yet continues planning without executing; zero tangible progress
• Root cause: Environment constraints (file write restrictions during flush) cause agent to stall in planning loop

#26 — Vault Agent (May 20)

• Scores: Accuracy 8, Efficiency 3, Tools 7, Skills 4, Overall 5
• Pattern: Runs memory promotion cron 14+ consecutive times with zero net-new content
• Root cause: No staleness detection in cron job; runs despite exhausted candidate pool

---

Root Cause Analysis

1. Model: deepseek-v4-flash:cloud is the common denominator

Agent	Model Used
All agents (trust, engineering, vault, market, health, etc.)	deepseek-v4-flash:cloud
Coach sidecar (critiquing agent)	deepseek-v4-flash:cloud (same model!)

Problem: deepseek-v4-flash:cloud is a fast/cheap model optimized for speed. It lacks the depth to:

• Detect when it's in a loop and break out
• Propose creative alternatives when blocked
• Self-assess whether it's making progress vs. spinning

The coach itself uses the same model to critique agents — meaning the model is grading itself.

2. No skill guardrails for loop detection

The agent config (openclaw.json) shows:

• timeoutSeconds: 600 (hard timeout only — no semantic loop detection)
• subagents.runTimeoutSeconds: 600
• No skills defined for any agent (skill lists are empty)

No agent has a "loop breaker" skill, a "staleness detector," or a "maximum-yield" guard.

3. Three specific failure patterns (all with the same fix direction)

Pattern	Agent(s)	Fix
Yield loop — blocked on human, keeps yielding	Trust	Add max yield counter + escalation
Planning loop — can't write files, replans instead of escalating	Engineering	Add env-constraint detection + file-write alternative path
Cron loop — runs job despite stale/no candidates	Vault	Add staleness detection before running cron

4. Coach architecture limitations

The sidecar (sidecar.mjs) at /home/sscoble/.openclaw/plugins/scobleclaw-agent-coach/sidecar.mjs:

• Has 48-hour dedup window — prevents spam but can mask patterns
• Uses temperature: 0.05 with response_format: json_object — produces consistent but shallow critiques
• The ISSUE_THRESHOLD = 5 means anything 5 or below gets reported; all 3 critiques hit exactly 5/10, which is suspiciously consistent

---

Proposed Improvements

Immediate (config/guardrails — can implement today)

1. Add loop-breaker skill to agent prompts

   • "After N consecutive identical actions (yield, no-op, re-plan), stop and escalate to human with context"
   • "When all ventures/tasks are blocked, generate summary report instead of looping"

2. Add cron staleness detection (already partially acknowledged in #26)

   • Skip promotion runs if no new entries since last successful run
   • Require minimum number of unique candidates before running

3. Add environmental constraint handling

   • When file writes fail, write to alternate path or queue instead of re-planning
   • Detect permission issues proactively

Medium-term (dispatch/coach improvements)

4. Use different model for agent vs. coach

   • Coach should use a stronger model (e.g. deepseek-v4-pro:cloud or deepseek-v3.2:cloud) for more nuanced critiques
   • Currently both agent and coach use deepseek-v4-flash:cloud — the critiquer is the same model being critiqued

5. Raise coach threshold or reduce dedup window

   • ISSUE_THRESHOLD of 5 with 48-hour dedup means we only see the worst cases. Consider threshold=6 with 24-hour dedup for earlier pattern detection.

6. Add "repetitive loop" as a specific tracked metric in coach critiques

   • Track consecutive identical actions, yield counts, and planning iterations

Long-term (architecture)

7. Consider model tiering — use deepseek-v4-flash:cloud for simple/known tasks, upgrade to deepseek-v4-pro:cloud or deepseek-v3.2:cloud for complex/problem-solving tasks
8. Add iteration budget — max N iterations per task before forced escalation
9. Coach should use a different judge model to avoid self-referential grading

---

Verdict

These critiques are VALID signals. They identify a real systemic issue: all agents share the same shallow model (deepseek-v4-flash:cloud) and lack loop-detection guardrails. The pattern is real — it's not coach noise.

The fixes are relatively tractable: add guardrail skills to agent configs, improve staleness detection, and tier model selection. The critical first step is adding "loop breaker" prompts and yield limits to agents.