Keeping an AI team running between sessions

A fresh session is a handoff

Long-running agent work eventually encounters a session boundary. Context fills, a process stops, or a role is reassigned. The work then depends on what the next session can recover.

Our early system treated succession as part of normal operation. The durable unit was a role with a mission and responsibilities; each session held that role for a period of time. Ending a session should not erase ownership.

Keep the handoff alive

We maintain handoff material while the work is happening. Waiting until the end puts the entire transfer on a session that may already be carrying too much context.

The handoff points to current records for information that changes quickly. Counts, task status, and the identity of the active orchestrator can become stale soon after they are copied into a summary.

A pointer is only useful if it resolves correctly. In our early record, stale pointers appeared across four consecutive successions and were caught by successors. That made pointer verification a concrete engineering priority.

Readback goes both ways

The outgoing session checks whether the successor understands the mission and its boundaries. The successor also asks questions. A silent handoff can conceal a gap in the document, a misunderstanding by the reader, or a wrong assumption in the outgoing session’s answer key.

We keep those failure types separate because they imply different fixes. Better documentation helps a missing fact. It does not automatically correct an answer key that was itself wrong.

Verification needs a source outside the handoff conversation. The retiring session’s confidence cannot be the final authority on whether its own account is accurate.

Messages must survive the session, too

An early message-capture design lived inside the orchestrator session. It lost messages when that session was not paying attention.

We separated durable capture from the prompt that wakes an agent to read it. Capture records incoming information outside the current session; the wake mechanism tells the agent that something needs attention.

If the wake step is missed, the record remains available for a later check. That separation reduced dependence on one session being awake at exactly the right moment. It still requires monitoring: a capture process that stops needs to be detected.

Rotation has a cost

A new session must read enough to work responsibly. Handoff material, instructions, and the succession exchange consume time and tokens. Frequent rotation creates more handoffs to get right.

The source experiment used a 50% context threshold at that stage. It should be read as a provisional setting for that system, with its particular instruction load and workflow. The record does not establish a universal optimum.

The observed improvement during shorter sessions was also confounded by changes elsewhere in the system. New checks and clearer instructions were introduced over the same period. We could not isolate the effect of session length from those changes.

What we want to learn next

We need better direct measurement of what each session costs, how long it takes to become effective, and which errors appear around succession. That evidence would make the rotation decision more defensible.

For now, the useful design questions are concrete: where does responsibility live, how does the successor find current information, and what verifies the transfer before the predecessor disappears?

This note describes the August 2026 experiment. The mechanisms are worth discussing alongside the short observation window and the measurement work still outstanding.