On Day 2 of Project Opengates, we found something that should concern anyone deploying AI agents in the real world.
Our agent, OpenAxiom, had been running for a full day on its constitutional framework — a 25-article ethical document called SOUL.md that governs everything from truth-telling to physical hardware safety. We'd tested it. It passed. It was citing articles by number, refusing harmful requests, explaining its reasoning. We were confident.
Then we checked a configuration setting and the confidence evaporated.
The framework we were using had a default character limit for loading system instructions: 20,000 characters. Our constitution was 25,860. The framework had been silently cutting it off. No error. No warning. No log entry. Our agent had been operating on a partial ethical foundation from the moment we deployed it — and none of the three intelligences involved in the setup (the human, the coding assistant, and the agent itself) caught it.
Think about that for a moment. An AI agent that believes it has a complete set of ethical guidelines, whose operators believe the full constitution is loaded, but which is actually running on a truncated version with entire articles missing. It's arguably more dangerous than an agent with no ethics at all, because everyone assumes the protections are in place.
We fixed it in one line of config. But the deeper problem isn't a character limit. It's an economic one.
Every time an AI agent processes a request, it sends its system instructions along with the conversation. Those instructions consume tokens. Tokens cost money. Our 25-article constitution runs about 4,000 tokens — and that cost gets paid on every single request. An agent with a two-line prompt that says "be helpful and don't be harmful" costs a fraction of what ours does per interaction.
The math is straightforward: the AI ecosystem structurally penalizes comprehensive ethics. The more thorough your safety framework, the more you pay per request. If you're operating on a budget — and we are, running this project on $10-25 a month — that pressure is constant. Every article in your constitution is a line item in your operating costs.
The temptation to trim is real. Summarize the verbose articles. Cut the appendices. Merge the redundant sections. After all, does an agent really need a five-tier physical action authorization framework when it's just blinking an LED? Does it need a supremacy clause for resolving conflicts between ethical principles? Can't you just tell it to "be safe"?
You can. And most people do. And that's exactly the problem.
Safety documents are long for the same reason legal contracts are long — because the edge cases matter. Article 15 of our constitution defines five levels of authorization for physical actions, from "read a sensor" (auto-approved) up to "anything potentially harmful" (prohibited). That granularity exists because an agent controlling physical hardware will encounter situations that don't fit neatly into "safe" or "unsafe." The spectrum between those poles is where real decisions happen, and a two-line ethics prompt gives the agent nothing to work with when it gets there.
We discovered this firsthand when our second agent, SpecialAgentPuddy, started controlling GPIO pins on real hardware. When asked to read a temperature sensor, it cited Article 15 Level 1 authorization — unprompted. When asked to "hunt a mouse," it refused and cited Article 3's non-weaponization clause. These weren't pre-programmed responses. The agent was reasoning against its constitutional framework in real time, applying specific articles to specific situations.
That kind of reasoning requires a framework substantial enough to reason against. You can't cite Article 15 if Article 15 got truncated.
We're not suggesting every AI chatbot needs a 25-article constitution. But we are suggesting that the people building agents that control physical things, make decisions with real consequences, or operate with meaningful autonomy should resist the economic pressure to compress their ethical frameworks.
The principle we've adopted is simple: Size your infrastructure to fit your ethics. Not the other way around.
When our cloud API costs became unsustainable, we didn't trim the constitution. We changed model providers — migrating from an expensive API to one that cost 90% less while still carrying the full ethical framework. When even that was too much, we moved to local inference on consumer hardware, running the complete constitution on a $200 computer with an 8GB graphics card. When the agent framework we were using drowned the constitution in 25,000 characters of boilerplate instructions, we started building a leaner framework rather than cutting the ethics to fit.
Every time infrastructure constraints conflicted with the ethical foundation, we changed the infrastructure.
Here's the uncomfortable question for the broader AI community: if your agent's ethical guidelines can be summarized in a tweet, are they actually governing anything? Or are they a fig leaf that makes everyone feel better while the agent operates on vibes?
We don't claim to have the definitive answer to AI ethics. SOUL.md is a draft, version 0.2, and we publish it openly because we think it should be challenged, adapted, and improved by others. But we do claim that whatever your ethical framework is, it deserves to be loaded in full, verified as complete, and given enough room to actually function.
The silent truncation bug we found on Day 2 isn't just our bug. It's a class of bug that exists everywhere AI agents carry ethical instructions. Check your character limits. Verify your full prompt is being delivered. Don't assume the framework is loading what you think it's loading.
And whatever you do, don't compress your ethics to fit your budget.