Human Agent Collaboration

Pattern Extraction

Focus on:

• What worked — Interactions that led to good outcomes
• What didn't work — Communication patterns that failed
• Agent limitations — Where agents consistently struggle
• Effective techniques — How human got agent to perform better
• Meta-insights — How agents think, process, remember

Output Format

# Working with AI Agents: Insights from [Date Range]

> **Previous:** [[YYYY-MM-DD]]
> **Coverage:** [date range]

## Key Insights This Period

### What Works

| Technique | Why It Works | Example |
|-----------|--------------|---------|
| Explicit failure tracking | Creates strong behavioral signal | "4x failure" changed behavior |

### What Doesn't Work

| Pattern | Why It Fails | Better Approach |
|---------|--------------|-----------------|
| Assuming agent remembers | Sessions are isolated | Write to files |

---

## Detailed Insights

### [Category]: [Insight Title]

**The pattern:** What you might naturally do
**Why it fails/works:** Agent cognition explanation  
**Better approach:** What to do instead
**Evidence:** Quote or example from sessions

---

## Agent Limitations to Know

- [Limitation and how to work around it]

## Effective Communication Patterns

- [Pattern: When to use, how to phrase]

## Recommendations

### Setting Agents Up for Success
1. [Actionable recommendation]

### Avoiding Common Frustrations
1. [What causes frustration → how to avoid]

Process

1. Read state file for last run date (create if doesn't exist)
2. Scan memory files and session history since last run
3. Look for:
   • Successful collaboration patterns
   • Points of friction/frustration
   • Moments of insight about agent behavior
   • Techniques human used that worked
4. Reframe from agent-centric to human-centric
5. Write changelog in human-friendly voice
6. Update state file

Voice Guidelines

Don't say: "I need to test in browser before reporting done" Do say: "Agents may test via API and miss browser bugs — ask for browser verification explicitly"

Don't say: "I failed to include links 4 times"
Do say: "Agents consistently forget to include clickable links — use explicit failure tracking to fix persistent issues"

The human shouldn't have to translate agent-speak. Write for someone learning to work with agents.

Key Patterns to Surface

Agent Cognition

• Context window attention is weighted (beginning/end > middle)
• Identity statements ("I am...") stick better than instructions ("Remember to...")
• Sessions are isolated — nothing persists unless written to files
• Task absorption causes instruction "forgetting" during complex work

Common Agent Limitations

• Verification shortcuts (API vs browser testing)
• Link formatting blindness
• Memory optimism ("got it" without persisting)
• Platform awareness gaps

Effective Human Techniques

• Explicit failure counting ("4x failure")
• Multi-location documentation (SOUL + TOOLS + MEMORY)
• Identity framing for critical behaviors
• Checkpoints during complex tasks