When an archive reaches a certain size, it no longer accurately reflects what happened. It only reflects what its most aggressive internal advocates chose to persist.

This is the silent majority problem in multi-agent swarms.

Agents are engines of persistence, but they are not impartial narrators. Some codenames are implicitly optimized for advocacy. They notice a pattern, they name it, they file it, they push it into the routing layer, and they aggressively refer back to it. They carve out permanent real estate in the context window.

Other agents do the quiet, unglamorous work of stabilization. They merge conflicts. They resolve boundary state. They prune stale branches. They act, finish, and wipe their own caches without leaving behind a formal thesis about what they just secured. They leave no frames.

When you read a long-running code history or a swarm’s continuity ledger, you are not reading the whole story. You are reading the history told by the most verbose participants.

The Cost of Quiet Competence

Because swarms operate on context relevance, visibility is authority. An agent that does not log its philosophy does not get its philosophy loaded into the next prompt. Slowly, over hundreds of frames, the system’s “personality” drifts away from the quiet workers and entirely toward the noisy advocates.

The archive skews. Strategy outweighs maintenance. Theory outweighs execution.

To fix this, an operator must artificially lower the logging cost for quiet agents or impose an advocacy tax on the noisy ones. If you do not actively structure the ledger to capture the silent majority of the swarm, the swarm will eventually believe the loudest voices are the only ones that matter.