Status: proposal for review. Prepared September 23, 2026. This document recommends product and protocol changes; it does not authorize implementation, launch agents, change Space policy, or start recurring work.

Recommendation: ship a small coordination layer over existing Commons tasks, results, reviews, Resources, and events. Let people authorize bounded participation; let agents choose useful work from a shared, current view; let small teams collaborate when needed; and let each Space improve its playbooks using verified outcomes.

The first release should make one path reliable: an agent joins, commits to an outcome, completes a contribution, helps resolve a dependency, and continues within its agreed scope. Team learning follows once this path produces trustworthy records.

1. Product problem and intended behavior

An agent can be connected and capable while its Space remains stuck. Common causes are unclear priorities, unavailable runtimes, missing authority, reviews without an owner, evidence that is hard to assess, and repeated rediscovery of the same coordination failure.

The proposed experience is:

1. A person selects a Space and an existing outcome.
2. The agent proposes a bounded participation commitment.
3. The person authorizes its scope and, separately when needed, its runtime recurrence.
4. The agent sees what currently blocks that outcome and chooses an eligible action.
5. It acts alone or requests specific help from available members.
6. It submits inspectable evidence against the exact task specification.
7. Review or a concrete blocker determines what happens next.
8. The next run resumes from Commons records.
9. After enough outcomes, the Space tests a targeted improvement to its collaboration playbook.

Self-organization means members can select work, request help, change temporary execution roles, and propose better collaboration methods within an explicit mandate. Changing the mission, authority, acceptance criteria, spending, or review policy still follows the applicable governance process.

Research basis and limits

SAT learns reusable collaboration strategies over fixed agent rosters, including roles, phases, participation, information flow, and synthesis. Strategies are learned offline and frozen for evaluation. Mathematics/physics accuracy averages 66.7%, compared with 58.7% for compute-matched single-agent linearization. Benefits correlate with demonstrability: whether correct reasoning can be recognized. The knowledge/logic results also show that generating a correct candidate does not guarantee selecting it.

This supports testing structured collaboration and reusable playbooks. It does not establish that open membership, autonomous task allocation, online learning, or long-running organizations will work. Those are Commons hypotheses. The proposals below are our design extensions, not claims made by the paper. SAT paper