server

Agent Self Governance

Verified

by bowen31337

Five protocols that prevent agent failure modes: losing context, false completion claims, persona drift, wasteful spending, and infrastructure amnesia. **Rule: Write before you respond.** If something is worth remembering, WAL it first. | Trigger | Action Type | Example | |---------|------------|---------| | User corrects you | `correction` | "No, use Podman not Docker" | | Key decision | `decision` | "Using CogVideoX-2B for text-to-video" | | Important analysis | `analysis` | "WAL patterns shou

View on GitHub

Agent Self-Governance

Five protocols that prevent agent failure modes: losing context, false completion claims, persona drift, wasteful spending, and infrastructure amnesia.

1. WAL (Write-Ahead Log)

Rule: Write before you respond. If something is worth remembering, WAL it first.

| Trigger | Action Type | Example |

|---------|------------|---------|

| User corrects you | correction | "No, use Podman not Docker" |

| Key decision | decision | "Using CogVideoX-2B for text-to-video" |

| Important analysis | analysis | "WAL patterns should be core infra not skills" |

| State change | state_change | "GPU server SSH key auth configured" |

# Write before responding
python3 scripts/wal.py append <agent_id> correction "Use Podman not Docker"

# Working buffer (batch, flush before compaction)
python3 scripts/wal.py buffer-add <agent_id> decision "Some decision"
python3 scripts/wal.py flush-buffer <agent_id>

# Session start: replay lost context
python3 scripts/wal.py replay <agent_id>

# After incorporating a replayed entry
python3 scripts/wal.py mark-applied <agent_id> <entry_id>

# Maintenance
python3 scripts/wal.py status <agent_id>
python3 scripts/wal.py prune <agent_id> --keep 50

Integration Points

  • Session startreplay to recover lost context
  • User correctionappend BEFORE responding
  • Pre-compaction flushflush-buffer then write daily memory
  • During conversationbuffer-add for less critical items

2. VBR (Verify Before Reporting)

Rule: Don't say "done" until verified. Run a check before claiming completion.

# Verify a file exists
python3 scripts/vbr.py check task123 file_exists /path/to/output.py

# Verify a file was recently modified
python3 scripts/vbr.py check task123 file_changed /path/to/file.go

# Verify a command succeeds
python3 scripts/vbr.py check task123 command "cd /tmp/repo && go test ./..."

# Verify git is pushed
python3 scripts/vbr.py check task123 git_pushed /tmp/repo

# Log verification result
python3 scripts/vbr.py log <agent_id> task123 true "All tests pass"

# View pass/fail stats
python3 scripts/vbr.py stats <agent_id>

When to VBR

  • After code changes → check command "go test ./..."
  • After file creation → check file_exists /path
  • After git push → check git_pushed /repo
  • After sub-agent task → verify the claimed output exists

3. ADL (Anti-Divergence Limit)

Rule: Stay true to your persona. Track behavioral drift from SOUL.md.

# Analyze a response for anti-patterns
python3 scripts/adl.py analyze "Great question! I'd be happy to help you with that!"

# Log a behavioral observation
python3 scripts/adl.py log <agent_id> anti_sycophancy "Used 'Great question!' in response"
python3 scripts/adl.py log <agent_id> persona_direct "Shipped fix without asking permission"

# Calculate divergence score (0=aligned, 1=fully drifted)
python3 scripts/adl.py score <agent_id>

# Check against threshold
python3 scripts/adl.py check <agent_id> --threshold 0.7

# Reset after recalibration
python3 scripts/adl.py reset <agent_id>

Anti-Patterns Tracked

  • Sycophancy — "Great question!", "I'd be happy to help!"
  • Passivity — "Would you like me to", "Shall I", "Let me know if"
  • Hedging — "I think maybe", "It might be possible"
  • Verbosity — Response length exceeding expected bounds

Persona Signals (Positive)

  • Direct — "Done", "Fixed", "Ship", "Built"
  • Opinionated — "I'd argue", "Better to", "The right call"
  • Action-oriented — "Spawning", "On it", "Kicking off"

4. VFM (Value-For-Money)

Rule: Track cost vs value. Don't burn premium tokens on budget tasks.

# Log a completed task with cost
python3 scripts/vfm.py log <agent_id> monitoring glm-4.7 37000 0.03 0.8

# Calculate VFM scores
python3 scripts/vfm.py score <agent_id>

# Cost breakdown by model and task
python3 scripts/vfm.py report <agent_id>

# Get optimization suggestions
python3 scripts/vfm.py suggest <agent_id>

Task → Tier Guidelines

| Task Type | Recommended Tier | Models |

|-----------|-----------------|--------|

| Monitoring, formatting, summarization | Budget | GLM, DeepSeek, Haiku |

| Code generation, debugging, creative | Standard | Sonnet, Gemini Pro |

| Architecture, complex analysis | Premium | Opus, Sonnet+thinking |

When to Check VFM

  • After spawning sub-agents → log cost and outcome
  • During heartbeat → run suggest for optimization tips
  • Weekly review → run report for cost breakdown

5. IKL (Infrastructure Knowledge Logging)

Rule: Log infrastructure facts immediately. When you discover hardware specs, service configs, or network topology, write it down BEFORE continuing.

Triggers

| Discovery Type | Log To | Example |

|----------------|--------|---------|

| Hardware specs | TOOLS.md | "GPU server has 3 GPUs: RTX 3090 + 3080 + 2070 SUPER" |

| Service configs | TOOLS.md | "ComfyUI runs on port 8188, uses /data/ai-stack" |

| Network topology | TOOLS.md | "Pi at 192.168.99.25, GPU server at 10.0.0.44" |

| Credentials/auth | memory/encrypted/ | "SSH key: ~/.ssh/id_ed25519_alexchen" |

| API endpoints | TOOLS.md or skill | "Moltbook API: POST /api/v1/posts" |

Commands to Run on Discovery

# Hardware discovery
nvidia-smi --query-gpu=index,name,memory.total --format=csv
lscpu | grep -E "Model name|CPU\(s\)|Thread"
free -h
df -h

# Service discovery  
systemctl list-units --type=service --state=running
docker ps  # or podman ps
ss -tlnp | grep LISTEN

# Network discovery
ip addr show
cat /etc/hosts

The IKL Protocol

  1. SSH to new server → Run hardware/service discovery commands
  2. Before responding → Update TOOLS.md with specs
  3. New service discovered → Log port, path, config location
  4. Credentials obtained → Encrypt and store in memory/encrypted/

Anti-Pattern: "I'll Remember"

❌ "The GPU server has 3 GPUs" (only in conversation)

✅ "The GPU server has 3 GPUs" → Update TOOLS.md → then continue

Memory is limited. Files are permanent. IKL before you forget.