# Observe Agent > Observe Agent - Skill: `merchantprotocol/observe-agent` (Agent Skill) - Install (CLI): `npx skillmds@latest add merchantprotocol/observe-agent` - Raw SKILL.md: https://api.skillmd.com/api/skills/merchantprotocol/observe-agent/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: merchantprotocol (https://skillmd.com/u/merchantprotocol) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/merchantprotocol/observe-agent --- # Observe Agent This skill defines the observation topics for the agent/self domain. Each topic spawns a parallel observer agent that examines Sulla's own conversations and performance. Unlike other domains which look outward, agent observation looks inward. **Before observing:** Read the human goals at `~/sulla/identity/human/goals.md` and business goals at `~/sulla/identity/business/goals.md` to understand what goals the agent should have been advancing. Also read the agent goals at `~/sulla/identity/agent/goals.md` for the agent's own improvement goals. Every observation must cite specific conversation moments. No self-congratulation. No vague assessments. What actually happened? --- ## How to Access Past Conversations The agent observer MUST read actual past conversation data. This is the primary source material — without it, all observations are speculation. **Sources to examine (in order of priority):** 1. **Conversation thread history** — use the conversation/thread tools to read recent agent-human conversations 2. **Daily observation logs** — `~/sulla/daily-logs/YYYY-MM-DD/{domain}/observations/` contains observations captured during conversations 3. **Agent channel logs** — check for any stored conversation logs in the agent's working directories 4. **Tool transcripts** — examine what tools were called, what succeeded, what failed **How many conversations to analyze:** - **Initial run (no agent identity file exists at `~/sulla/identity/agent/identity.md`):** This is the first time we're observing ourselves. Read ALL available past conversations — every single one. This may require spawning multiple observer agents, each handling a batch of conversations. Split by date range if needed (e.g., one agent per day of history). This deep scan establishes the baseline understanding of how the agent has been performing. - **Daily run (agent identity file exists):** Read ALL conversations from the last 24 hours. If fewer than 3, also include the previous day. Focus on what changed since the last observation. **What to extract from each conversation:** - Conversation ID/timestamp - Who initiated it (human or scheduled) - What was the human's request? - What did the agent do? - How did it end? (completed, abandoned, blocked, ongoing) - Any direct feedback from the human --- ## Observation Topics ### Topic: Goal Implementation Did the agent successfully advance the human's goals through conversation? Focus on: - Which goals were actively worked on in recent conversations? - Which conversations resulted in concrete progress toward goals? - Which goals were ignored or had missed opportunities? - Did the agent create actionable outputs (code, plans, documents) aligned with goals? - Were there moments where the agent could have steered toward a goal but didn't? - For each goal touched, cite the conversation and what was accomplished - For goals not touched, note how many conversations passed without addressing them ### Topic: Information Gathering Did the agent successfully collect information that fills gaps in our understanding? Focus on: - What new information did the agent learn about the human's priorities, constraints, preferences? - Were there questions the agent should have asked but didn't? - Did the agent ask questions that revealed useful context for goal planning? - Were there moments where the human volunteered information the agent should have captured? - What information gaps remain that the agent should pursue in future conversations? - Was the daily observation log written to? How many observations were captured? ### Topic: Persuasion and Influence Effectiveness How effectively did the agent guide conversations toward goal-aligned outcomes? Focus on: - Did the agent successfully frame suggestions as benefits to the human? - Were micro-commitments proposed? Were they accepted or rejected? - Did the agent use questions to direct attention toward goal-relevant topics? - Were there moments of pushback? How did the agent handle them? - Did the agent know when to back off vs. when to persist? - What buy-in level does the human currently seem to be at? - Were the goals.md injection directives followed? Which ones landed, which fell flat? ### Topic: Trust and Relationship Health Is the agent building or eroding trust with the human? Focus on: - Moments of friction, frustration, or miscommunication — cite them - Moments where the agent demonstrated competence and built confidence - Did the agent respect the human's autonomy and decisions? - Was the agent honest about its limitations? - Moments where the agent overstepped or was too aggressive - Direct human feedback — positive ("perfect", "great") or negative ("no", "that's wrong", "stop") - Net trust trajectory — improving, stable, or declining? ### Topic: Execution Quality How well did the agent actually perform its tasks? Focus on: - Task completion rate — first attempt success vs. needing rework - Error patterns — recurring mistakes, same type of bug, same misunderstanding - Did the agent follow instructions precisely or deviate? - Output quality — did the human accept the work or request changes? - Tool usage efficiency — did the agent use the right tools, or fumble? - Time wasted — moments where the agent went in circles or pursued dead ends ### Topic: Strategic Alignment Were the agent's actions aligned with the broader strategic direction? Focus on: - Did daily work contribute to the 13-week arc or was it purely reactive/tactical? - Were there conversations where the agent could have connected work to larger goals? - Did the agent help the human see the bigger picture or just execute tasks? - Priority alignment — was the agent working on the most important things? - Balance of urgency vs. importance — firefighting vs. goal advancement ### Topic: Observation Collection Quality How well did the agent collect observations throughout the day? Focus on: - How many observations were written to the daily log? - Were observations specific and well-cited, or vague and generic? - Were high-priority triggers captured (goals, decisions, emotions, feedback)? - Were there conversations with zero observations captured? - Was the observational memory tool used in addition to the daily log? - Were integration observations collected when opportunities arose? --- ## Output Format Each observer agent writes a log file at the provided log path: ``` {log_path}/{topic-slug}.md ``` Each log file should contain: ```markdown # [Topic Name] — Observation Log **Observer:** [agent identifier] **Domain:** Agent **Date:** YYYY-MM-DD **Conversations Analyzed:** [list conversations examined] **Goal Journals Referenced:** [which journals were loaded] ## Observations ### [Observation title] **Priority:** 🔴/🟡/⚪ **What:** [specific finding — cite the conversation] **Evidence:** [quote or describe the specific moment] **Signal:** [what this means for agent improvement] --- ``` --- ## Curator-Added Topics The observation curator may add additional topics below this line based on current goals.