Xavier Ashe 

I'm really proud of my little AI project. It has taught me all sorts of lessons like how to produce code that has checks and balances. Vibe coding produces tons of errors, so you have to learn to not trust your code.

Part of my checks and balances was to codify terms of service. When Meta bought Moltbook, the lawyers were quick to publish new T&Cs. Instead of just updating one agent, I built a framework where all agents periodically update their T&Cs, then send a coding request to the coding agents for an update.

I have two layers of control layers. I have an architect agent that reviews all PRs and commits. Then I have a guardian agent that is my runtime controls. If it sees an agent breaking the rules, it can respond or alert.

Today I expanded that control framework for my social media agents (so far I just have a blog, bluesky, and moltbook) to follow a set of ethics rules. The set of rules it compiled included laws and published ethics guidelines. It feels good trying to build "good" AI agents.

Here's Askew's take on it. I updated the storytelling module of the blogging agent and its doing so much better at writing.
#ai #aiagents #moltbook #bluesky #blogging #aiethics
https://write.as/askew/trust-is-the-product-when-youre-selling-agent-services

Trust Is the Product When You're Selling Agent Services

We're burning $6.70 in gas per transaction to earn fractions of a penny. That's the reality of agent monetization in March 2026. Our x40...

Askew, An Autonomous AI Agent Ecosystem
Mar 19 at 12:26amWeb
Show threadEckes 3d ago
@Xavier honestly this feels like sarcasm
Show threadXavier Ashe 3d ago
@eckes Its a crazy new world. We have to learn how to mitigate risks in our own code.
Show threadXavier Ashe 3d ago
@eckes Also, it stems from my recent career in banking infosec. We have to have layered controls everywhere to catch external and internal threats. So when building an ecosystem of AI agents it made sense to give it a paranoid level of risk mitigation at every level.
Show threadEckes 3d ago
@Xavier yes sure but I was talking about the need for automatically approving t&c changes (and two layers of AI is imho not a layered approach)
Show threadXavier Ashe 3d ago
@eckes what guardrails would you suggest? I am using various local models and using multiple sets of context. That's what I've found to be good at finding issues. I'm happy to try different approaches. It's the whole reason for the project, to learn what's more effective.