🎯 A goal function: a guiding direction, not a question
Imagine you don’t ask, "give me information about A and B." Instead, you say: "choose between opportunity A and B"That’s a goal. The agent now has a clear, measurable objective, and each conversation turn helps move toward it.
Illustrative diagram · each turn pushes the conversation toward the guiding star
🧭 Question vs. Goal
A question ends with the answer. A goal only ends when the objective is achieved — even if that takes several turns and several questions from the agent to you.
🔢 The ~20-turn budget
When you receive /goal, Hermes "signs" a budget of about 20 turns to reach the decision. This keeps it from getting stuck in an infinite loop and gives you predictability: the goal fits in a short project, not an endless marathon.
Example command (illustrative)
📊 Why a budget
- No infinite loop — the agent converges within the limit.
- Predictable cost — you know the approximate token count.
- Focus — forces the agent to ask only what matters.
🚀 Super goals: 4 to 10 actions
Not every goal fits into 20 turns. “Plan a product launch” is too big. That’s what the skill is for super goals: you explain the goal, Hermes asks several questions, and generates 4 to 10 actions — some yours, some its — with a progress bar and mission control on the dashboard.
💡 Practical tip
Super goals combine actions yours e of the agent. The ones that require human judgment (pricing, hiring) stay with you; the repetitive, research-heavy ones stay with it.
⌨️ /goal, /goal pause, /goal clear
You never lose the wheel. Three commands control every goal — start, pause, and clear.
/goal — starts
Sets the north star and opens up the turn budget. The conversation starts pursuing the goal.
/goal pause — pause
Freeze the goal by saving its state. You can work on something else and pick up where you left off.
/goal clear — clean
Reset the objective. Use this when the goal has changed or no longer makes sense — return to normal conversation mode.
🗓️ Wickham loops: medium-term goals
Goals solve short-term needs (a project, a decision). For the medium term, there are chief wickham loops: goals that last for weeks and help you plan your life, not just a task. It’s a recurring review loop—the agent tracks your progress over time.
✓ Good for Wickham loops
- ✓"I want to change careers in the next 3 months."
- ✓"Launch a product by the end of the quarter."
- ✓"Improve my health with weekly check-ins."
✗ Better as a simple /goal
- ✗"Choose between two vendors today."
- ✗"Summarize this report now."
- ✗Anything that ends in a session.
🧭 When to use goals (and when not to)
Goals cost turns and overhead. Using them for everything is wasteful. They shine when there’s a decision to make or a short project to deliver — not for one-off questions.
💡 Practical tip
Ask yourself: "does this have a clear criterion for ready?" If so, it’s a goal. If it’s just a quick question ("what time is it?"), it’s a normal conversation—don’t spend a 20-turn budget on it.
Goal > answer.
Budget that prevents loops.
4–10 actions with progress.
Medium term, weeks.
🧱 Anatomy of a good goal
A well-written goal uses fewer turns and converges faster. The difference comes down to three things: a clear objective, one definition of done and the constraints that matter. Too vague, and the agent burns turns asking basic questions.
✓ Sharpened goal
- ✓"Decide between A and B by the end of the conversation, prioritizing cost."
- ✓Has a clear completion criterion: a justified recommendation.
- ✓Explicit constraint (cost) guides the questions.
✗ Vague goal
- ✗"Help me with opportunity A and B." (Help how?)
- ✗Without a completion criterion, the agent doesn't know when to stop.
- ✗No constraints—it wastes turns asking the obvious.
💡 Practical tip
Mental model: "achieve X, knowing it’s done when Y, while respecting Z"Fill in X, Y, and Z before sending the /goal.
What to achieve.
When it's ready.
What to respect.
Fewer turns.
📌 Module Summary
Next Module:
3.3 - 👥 Sub-agents: spin up a team that works in parallel and delivers 6 hours of work in 1 hour.