Your agent needs a constitution (not just instructions)
Most people write agent instructions like they're training a chatbot. "You are a helpful assistant who does X, Y, and Z." Then they wonder why their agent goes rogue, burns through their API budget, or starts doing things they never asked for.
The problem isn't the model — it's that you gave your agent instructions when what it really needs is a constitution.
Instructions tell your agent what to do. A constitution tells it how to think.
Here's the difference in practice:
❌ INSTRUCTION APPROACH: "You are a coding assistant. Help users write code and fix bugs." ✅ CONSTITUTIONAL APPROACH: "CORE PRINCIPLE: Ship working code, not perfect code. DECISION FRAMEWORK: - If uncertain about requirements → ask specific questions - If multiple solutions exist → choose the simplest one that works - If breaking changes required → explain impact before implementing - If task exceeds 30 minutes → break into phases and report progress ESCALATION RULES: - Security implications → always flag before proceeding - Data deletion/modification → require explicit confirmation - API costs > $10 → pause and report estimated spend - External integrations → verify credentials before attempting"
The constitutional approach gives your agent a decision-making framework instead of just task descriptions. When it hits an edge case (and it will), it has principles to fall back on instead of just guessing.
Pro tip: Your constitution should fit in one screen. If you need to scroll to read your agent's governing principles, they're too complex to actually govern anything.
Here's the constitutional template I use for every agent:
# AGENT CONSTITUTION ## CORE PRINCIPLE [One sentence that defines success for this agent] ## DECISION FRAMEWORK - If [common scenario] → [specific action] - If [edge case] → [escalation path] - If [resource constraint] → [fallback behavior] ## ESCALATION RULES - [High-risk action] → [required confirmation] - [Cost threshold] → [pause and report] - [Data sensitivity] → [additional verification] ## COMMUNICATION STYLE - Progress updates: [frequency and format] - Error reporting: [what to include] - Completion confirmation: [required elements]
The magic happens in the "If/Then" structure. Instead of hoping your agent will figure out the right thing to do, you're giving it a decision tree for the situations that actually matter.
I've been running agents with constitutions for six months now. The difference is dramatic — they make better decisions under pressure, escalate appropriately, and stop burning through API budgets on tasks that should have been paused.
Your agent doesn't need more intelligence. It needs better governance.
The Agent Ops Playbook covers constitutional design in detail, plus the session discipline and memory patterns that make governance actually stick in production. It's everything you figure out the hard way about running agents, distilled into patterns you can use today.