Agent autonomy is not the same as good collaboration

Every AI collaboration tool I've used is fundamentally turn-based. You ask, it answers, the conversation ends, state resets. Even when the AI reads all your materials, you still have to wait for it to finish. Your modifications only exist as subsequent messages. The AI has no reliable way to know which of its earlier judgments are still valid after you've changed something upstream.

Chat is good for exchanging messages. It's terrible for expressing a task environment that both sides can continuously modify.

I built something called SharedGround to test what happens when you remove the turn-based constraint. The core idea: human and agent share a workspace where both can observe, decide, and act without waiting. Sources, evidence, claims, briefs — all live as structured objects with version numbers. The agent runs in a continuous Observe-Decide-Act-Feedback loop.

The first version worked. The demo loop ran. And then I discovered a critical flaw: the API was returning a complete new state that replaced the browser state entirely. If I edited a claim while the agent was running, the old API response — generated from a snapshot taken before my edit — could silently overwrite my changes. Human edits were vulnerable to race conditions baked into the architecture.

V0.2 inverted the ownership model. The API no longer returns state — it returns only proposed actions. Each action carries an expectedVersion. The browser reducer checks: has this object changed since the agent last saw it? If yes, reject. Log the rejection with a reason. The human always wins, enforced at the reducer level, not in a prompt.

This shift clarified something I'd been circling: the goal of human-agent collaboration isn't making the agent more autonomous. It's designing how state is shared and how control is exchanged. Autonomy without clear boundaries is just a polite way of saying the system can overwrite your work.

The permission matrix in SharedGround is deliberately asymmetric. The agent can propose claims, challenge claims, edit briefs. It cannot confirm, finalize, or mark anything as human-reviewed. Only the human can close a claim or finish the task. Every boundary is structural, not rhetorical. The system has 125 deterministic tests verifying each one.

I don't think AI collaboration tools need to be less capable. I think they need clearer contracts about who owns what and when. Structured actions, version-checked writes, and reducer-enforced permissions aren't exciting features. But they're what separates a tool that helps you think from one that silently replaces your thinking.