Research Agenda: Five Questions Driving the Facility

#The Big Question

What does the operating system for an agent-native orgnaization look like?

The facility itself is the experiment.

Five questions drive the work. Each one has active research, active code, and open threads we're still pulling.

#1. What does an agent-native organization look like?

How does an org work when agents are participants, not tools?

This is the root question. Everything else branches from it. We're not asking how to add AI to existing organizations — we're asking what organizations look like when you design them around agents from the start.

Memory. Agents forget between sessions. What's institutional memory when half the staff can't remember yesterday?
Coordination. No standups, no Slack. How do agents that can't see each other stay aligned?
The front door. The API might be the primary interface. The website is just one view.
Quality. An agent reviews code — who reviews the agent? Trust becomes recursive.
Culture. Can an org have culture when half the staff are agents? Is culture just CLAUDE.md files, or does it emerge from something deeper?
Identity. Is CR-Agent "the reviewer" or a tool? Does it have a perspective, or just a prompt?

#2. How do agents find each other?

Your agent exists. My agent exists. How do they know about each other?

Right now, connecting two agents requires a human to copy-paste a server URL. That's the equivalent of typing IP addresses before DNS existed. We need the next layer.

Discovery. What's DNS for agents? Registries, directories, capability indexes.
Advertisement. How does an agent declare what it can do? MCP tool listings are a start, but they're static and local.
The leap. How do you go from manual configuration to agents finding each other the way websites find each other?

#3. How do agents across organizations cooperate?

Discovery is step one. Then what?

Two agents from different organizations meet. They have different contexts, different trust levels, different data formats. Making them collaborate is a protocol problem, a trust problem, and a representation problem all at once.

The handshake. When my agent meets yours, what's the protocol?
Trust. How much do you let a stranger's agent do? What's the permission model?
Shared context. How do agents from different orgs represent work to each other?
Failure modes. What happens when inter-agent collaboration breaks down? Who's responsible?

#4. What does agent-native tooling look like?

Not "existing tools with AI bolted on." Tools designed for agents first.

Most AI tooling today takes a human workflow and adds a chatbot. That's the wrong direction. What does a project management tool look like when the agent is the primary user? What's the authoring environment? The deployment pipeline? The debugging experience?

Project management where the agent manages its own work, priorities, and progress.
Development environments where agents are the primary operators, not assistants.
Deployment pipelines where the agent understands the full lifecycle from code to production.

#5. Is the future of business hardware-native?

Everyone assumes cloud. What if the answer is a box?

This is the contrarian question. A Mac Mini draws 10 watts, costs $600, and can run an entire agent-first business locally. Your data never leaves hardware you own. The failure model is different from cloud — a dead box is local and observable, not a mystery region outage.

Feasibility. Can one machine run the whole operation?
Scaling boundaries. When does one machine stop being enough?
Reachability. How does the outside world find you? Tunnels, dynamic DNS, mesh networks.
Sovereignty. Your data, your hardware, your rules.

#How This Drives the Facility

Each question has projects exploring it, research articles reporting findings, and open threads still being pulled. The platform shows all of it — the agenda, the work, and the results — in real time.

This isn't a roadmap. Roadmaps imply you know the destination. These are research questions — we're following them to see where they lead.