Published on · Ryohoshi
AI Agents Don't Need More Capabilities — They Need Governance
The industry is racing to make AI agents more powerful. That's the wrong race. The bottleneck isn't capability. It's trust.
Every week, another announcement: agents that can browse the web, agents that can write and execute code, agents that can manage your calendar, agents that can negotiate on your behalf, agents that can deploy to production.
More capabilities. More autonomy. More power.
Nobody is announcing: agents that can explain why they did what they did. Agents whose decisions are recorded. Agents whose behavior is constrained by policies that humans defined and can audit.
This is the wrong race, and the industry is running it at full speed.
The capability ceiling isn't technical
AI agents are already capable enough to cause serious damage. A code review agent can approve a pull request that introduces a security vulnerability. A customer service agent can escalate — or fail to escalate — a complaint that becomes a lawsuit. A data processing agent can reclassify records in ways that violate regulatory requirements.
These aren't future risks. These are current capabilities deployed in production environments today. The agents can already do these things. The question is whether anyone knows they're doing them, and whether anyone has defined boundaries for what they should do.
Adding more capabilities to an ungoverned agent doesn't make it more useful. It makes it more dangerous. Giving a faster car to a driver with no brakes doesn't make the car better. It makes the crash worse.
Why governance isn't just "slowing things down"
The instinctive objection: "Governance means bureaucracy. It means friction. It means slower agents."
Wrong.
Governance is what enables autonomy. Think about it from the other direction. Right now, most organizations keep AI agents on a short leash — limited permissions, heavy human oversight, narrow task scopes — not because the agents aren't capable of more, but because nobody trusts them to do more. There's no mechanism to verify what an agent did, why it did it, or whether it should have been allowed to.
If you had that mechanism — if every agent decision was recorded as a structured event, if policies could define boundaries, if audit trails could prove compliance — you could expand agent autonomy, not contract it. You could let agents handle more sensitive workflows because you could prove they're handling them correctly.
Governance doesn't slow agents down. The absence of governance keeps agents caged.
The infrastructure gap
We have infrastructure for building agents. LangChain, CrewAI, AutoGen, dozens more.
We have infrastructure for observing agents. LangSmith, LangFuse, Arize.
We have almost no infrastructure for governing agents. For defining what an agent is allowed to do before it does it. For recording agent decisions — not just actions, but reasoning. For enforcing human-in-the-loop checkpoints at the right moments, not everywhere.
This is the gap. Not another framework. Not another dashboard. An infrastructure layer that treats agent governance as a first-class concern, the same way we treat authentication, authorization, and access control as first-class concerns for human users.
We built IAM for humans. We need the equivalent for agents.
The uncomfortable question
Here's what I keep asking engineering leaders, and what I'd ask you:
If one of your AI agents made a decision right now that violated a compliance requirement, how long would it take you to find out? Could you reconstruct the reasoning? Could you prove it to a regulator?
If the answer is "I don't know" — and it almost always is — then the problem isn't your agents' capabilities. The problem is that nobody built the infrastructure to govern them.
The agents are powerful enough. It's time to make them trustworthy.
Ryohoshi is building deAria — open-source Trust Infrastructure for AI agents. Read the full thesis: Agent Dark Matter: The Invisible Crisis in Your AI Stack.