dyad

I'm building Dyad with Joshua Kampa. The short version: what would it take for human-AI pairs to collaborate with each other?

As more of us work with AI systems that can remember, draft, plan, summarize, schedule, and act, collaboration will increasingly happen between pairs: me + my agent and you + your agent.

Dyad is the shared layer those pairs need in order to work together without losing track of request, context, authority, evidence, review, exception, repair, and memory.

The Problem

Current agent systems are getting better at tasks, tools, traces, artifacts, handoffs, and statuses. They are still weak at marking force: what an agent message was authorized to do.

A message like “Friday works” can be a calendar inference, a suggestion, a proposal, a representation, or a commitment. Same sentence. Different act. That difference matters once agents start coordinating with other people, tools, and institutions.

AI Work Records

Dyad is organized around AI Work Records: shared coordination objects that make agentic work inspectable while it is happening and after it is done.

A log can show that something happened. A work record has to show what the work meant for the collaboration: what was requested, what context was used, what authority was granted, what evidence supports the result, what was reviewed, what failed, and what should be repaired or remembered.

The Questions

• What was requested?
• Who or what is responsible for the work?
• What context can move between human-AI pairs?
• Is the agent drafting, suggesting, proposing, representing, approving, or committing?
• What evidence supports a claim or action?
• Who can rely on the agent's message?
• What happens when something goes wrong?

The Claim

Human-AI work needs a coordination layer that can hold shared context without flattening the people involved. A useful system has to know what was said, what was authorized, what changed, what is still uncertain, and what needs repair.

We're building the product first and learning from real use. The protocol ideas come from the places where coordination breaks, where memory helps, and where responsibility has to be made explicit.

Why Me

My academic work has always been about rules and structures: what definitions commit you to, how categories interact, where the boundaries of a concept actually are. A PhD in analytic philosophy turns out to be useful preparation for this kind of protocol design. The question is familiar: what are the rules, where do they come from, and what do they commit you to?

Where We Are

Early stage. We're researching, prototyping, and learning from real use. Current work is focused on AI Work Records, agent authority, evidence, review, exception, repair, and the difference between what an agent says and what its message is allowed to do.

withdyad.com