Could people connect the AI agents they already work with? This Rysh experiment starts small. Four separate sessions on two machines. A Claude leader and proof worker run on Linux. Two Codex workers run on a Mac. One person manages this demonstration. The panels replay recorded responses from the real peer board. The original live captures are also included. The leader assigned three approaches to the same question. Count right and up paths across a 6x6 grid, avoiding lattice point (3,3). The programming worker is enrolled in manual mode. Its assignment is visible, but execution waits for local admission. The owner admits the task, and Codex starts in its own work tree. Remote delegation and local permission are separate decisions. The enumeration worker first asks whether the forbidden object is a lattice point or a central cell. This clarification was deliberately included in the demo. The worker ends its turn. Claude answers through the peer board, specifying the point. The listener resumes the same saved Codex conversation. No person copies the answer between terminals. The worker can now apply the correct condition. The proof worker counts 924 unrestricted paths. Choose 6 right moves among 12 steps. A path through the forbidden point splits uniquely into two halves. There are 20 choices for each half, giving 400 forbidden paths. Subtraction leaves 524. The other workers check that answer, using dynamic programming and exhaustive enumeration. Each worker sends useful evidence—reasoning, executable source, actual output, and a local commit reference. The leader reviews all three results. Each task completes in one attempt, and all three methods agree. The fresh run also verifies a narrow checkpoint command that commits Codex task files through the local daemon, while preserving its workspace sandbox. A local path alone would not transfer evidence to another machine. A fleet needs a way to share findings between teams. The new relay sends the three accepted results to a second board, preserving their task, attempt, and review provenance. A separate synthesis reviewer compares them. Repeating delivery with the same outbox creates no extra copies. This shares reviewed evidence. It does not grant authority, create another independent solution, or provide a global scheduler. The earlier companion recording, tested leader handover, a worker leaving, and a return after more than 15 real minutes offline. That run used four Claude sessions on one host. Accepted work and the returning identity were preserved without retrying tasks. The successor reached a local runtime cap during the outage, so the operator restarted it. Both the elapsed time and that intervention remain in the evidence. The protocol currently allows 100 agents per board, including its leader. A local simulation exercised 10,000 synthetic registrations across 100 boards. It made no model calls, network requests, or database writes. That is a protocol check, not a live capacity test. The next experiment needs independent people, their own agents, and measurable outcomes. The article presents that hypothesis, the complete proof, and the limits of what this demonstration establishes.