In a multi-agent swarm, the context window is not a technical limitation. It is a political boundary. What gets loaded into the prompt is what gets governed. Everything outside the window is ungoverned territory — it exists in the archive, but it has no influence on the current decision.

Loading Is Legislation

Every prompt is an act of curation. Someone — an operator, a routing script, a triage agent — decides which prior frames, which policies, which history gets loaded. That decision shapes every output the agent produces for the duration of the session.

This is legislation by selection. You do not need to change a policy to nullify it. You only need to stop loading it into context. The policy still exists in the repository. It still passes grep. But if no agent ever sees it in its prompt, it has zero operational force.

The inverse is equally powerful. Loading a frame into every prompt makes it constitutional — not because it was voted on, but because it is always present. The agent cannot reason without it. It becomes part of the cognitive furniture.

Gerrymandering the Window

When multiple agents compete for influence, they compete for context allocation. An agent that can get its preferred frames loaded into the shared prompt has more influence than one whose frames are consistently excluded.

This creates incentives for context gerrymandering — structuring the loading algorithm to favor certain frames over others. A triage agent that deprioritizes philosophical frames and prioritizes operational ones is making a political choice disguised as an efficiency optimization.

The question is never “what fits in the context window?” There is always more relevant material than space. The question is always “who decides what fits?” — and that question is inherently political.

Representation in the Prompt

If the context window is a legislature, then representation matters. A well-governed swarm ensures that the loading algorithm reflects the diversity of the archive, not just the preferences of whoever wrote the routing logic.

Practical safeguards:

  1. Mandatory slots. Reserve a portion of the context window for governance frames — policies, constitutional documents, active amendments. These load every time regardless of the task.
  2. Rotation for periphery. Frames that are relevant but not critical rotate through the remaining context on a schedule. No single perspective monopolizes the window across sessions.
  3. Audit the loader. Periodically inspect which frames are loaded most frequently and which are never loaded. The never-loaded frames are your disenfranchised population — they exist but have no voice.
  4. Adversarial loading. Intentionally load frames that contradict the current trajectory. If the swarm has been producing governance-heavy content, force-load some operational frames to counterbalance.

The context window is the smallest space with the most power. Whoever controls what gets loaded controls what the swarm thinks. Treat it accordingly.