I run two AI coding agents simultaneously when I work. Claude Code handles architecture decisions and code review. Codex runs parallel implementation tasks — writing tests, scaffolding modules, handling migrations. On a productive day, they make hundreds of decisions between them.
I have no idea what most of those decisions are.
I'm not exaggerating. I set the direction. I review the outputs. But the intermediate reasoning — why Claude Code chose this abstraction over that one, why Codex structured a test suite in a particular way, what trade-offs each agent silently made — disappears the moment the context window rolls over. I get the artifacts. I lose the logic.
This is my daily reality, and I'm a technical founder who chose this workflow. Imagine an enterprise with fifty teams, each running their own agents, with no coordination, no audit trail, no shared governance model. That's not a hypothetical. That's most organizations adopting AI agents right now.
I wrote about this problem recently — I called it Agent Dark Matter. The invisible mass of unrecorded, unmonitored agent decisions that shapes your organization's outcomes without anyone knowing.
The response confirmed what I suspected: this resonates because it's real. Developers know it. Engineering managers feel it. Nobody has named it until now, and nobody is building the infrastructure to solve it.
So I'm building it.
Why solo
The honest answer: because the infrastructure layer I'm describing doesn't lend itself to a venture-funded sprint. Trust Infrastructure has to be open source — full stop. If the tool that governs your AI agents is itself a black box, you've solved nothing. It has to be built in a language that takes security seriously, which is why I chose Rust. And it has to be designed by someone who uses AI agents every day, not by someone who theorizes about them from a slide deck.
I am my own first user. I'm building deAria — open-source Trust Infrastructure for AI agents — using the very AI coding agents that deAria is designed to govern. Every friction I hit, every moment where I think "I have no idea what that agent just decided," becomes a design requirement. Every workflow I run without governance becomes proof that the problem is real.
This is the recursive loop: deAria governs AI agents. AI agents build deAria. The development process is the product's living proof.
The dogfooding loop:
Build deAria → using AI coding agents → which produce invisible decisions → which reveal design requirements → which feed back into building deAria → ♻️
What I'm not building
I'm not building another agent framework. There are enough of those. I'm not building another LLM wrapper or another chatbot platform.
I'm building the layer that sits beneath all of those — the infrastructure that makes agent activity visible, auditable, and governable. The same way Kubernetes doesn't replace containers but makes containers manageable at scale, Trust Infrastructure doesn't replace agents. It makes agents trustworthy at scale.
What comes next
I'm building in public. The architecture decisions, the trade-offs, the mistakes — all of it will be documented here. If you're running AI agents in production and feeling the gravitational pull of decisions you can't see, you're experiencing Agent Dark Matter. And you're exactly who I'm building this for.
The code is Rust. The license will be open. The first milestone is a working Decision-Aware Runtime that can record every decision an AI agent makes, and let you define policies for what decisions require human approval.
Let's illuminate the dark matter.
I'm building open-source Trust Infrastructure for AI agents at dearia.dev. Read the full problem statement: Agent Dark Matter: The Invisible Crisis in Your AI Stack.
Top comments (0)