Why does a routing supervisor's context balloon by turn 15, and how do you fix it?

## Why the window grows A supervisor topology has one property that is simultaneously its strength and its scaling limit: everything passes through the centre. Specialist A answers, the answer lands in the supervisor's context so it can decide what is next; specialist B answers, that lands too. Nothing is ever removed. By turn 15 of a moderately chatty run the router is carrying the accumulated output of a dozen specialist turns, most of which mattered only for the one decision that followed them. A concrete shape: a city 311 supervisor consults potholes, permits, sanitation and housing agents over a complex mixed-use complaint. Each returns a few thousand tokens of detail because nobody defined what it should return. The supervisor's own context passes 30k tokens by turn 15 - and it is still only choosing a name from a list of four. ## Why it hurts more than the token bill **Cost is quadratic-ish across a run.** Each routing decision re-sends the whole accumulated context. Turn 15 costs roughly fifteen times turn 1 for a decision of identical difficulty. Prompt caching softens the repeated prefix but does not remove the growth, since every turn appends new tail content. **Latency follows.** Time to the routing decision grows with input length, and the supervisor sits on the critical path of every hop. **Quality degrades - this is the part candidates miss.** Long contexts suffer what practitioners call context rot: retrieval and instruction-following weaken as the window fills. A degrading router forgets its completion criteria, re-dispatches an agent that already answered, or starts answering the user itself. The failure looks like a bad model when it is a bad context budget. **And then the run stops.** Window exhaustion turns a slow system into a broken one, usually mid-task. ## The return contract The highest-leverage fix is defining, per specialist, what it is allowed to hand back. A good return contract is a small structured object: a one-paragraph outcome, a status, any facts the router genuinely needs for the next decision, and references to anything large. The specialist's internal reasoning, tool calls and intermediate results stay in its own context and die with it. The test is simple: the router needs enough to decide *who acts next and whether we are done*. It does not need the specialist's working notes. If your summaries are being written for the end user rather than for the router, they are the wrong artifact - produce the user-facing text once, at the end, from stored artifacts. ## The task ledger Even with tight returns, raw history accumulates. The stronger move is to stop feeding history at all and feed a maintained ledger instead: the goal verbatim, the completed steps with one-line outcomes, the open questions, and the completion criteria. The supervisor's prompt then grows with the *number of steps*, not with their verbosity, and the goal stays near the top of the window where it survives. This also makes the supervisor closer to stateless per decision, which has a pleasant operational side effect: you can replay a routing decision from its ledger alone when debugging. ## Externalize the bulk Large outputs - a permit history, a full inspection report, a generated document - belong in a store with a handle, not in the conversation. The supervisor routes on the handle plus a description; whichever specialist actually needs the content fetches it. This is the same discipline as passing a file path instead of a file. ## Compaction as the backstop When the window still fills on genuinely long runs, compact: summarize the older middle of the transcript while preserving the goal, the ledger and the most recent turns verbatim. Compaction is a backstop, not a strategy - if you are compacting every few turns, your return contracts are too loose. ## Knowing it is happening Instrument tokens-in per routing decision as a time series across a run, not just totals per run. A rising line with a flat decision difficulty is the signature. Watch alongside it: hops per run, repeat dispatches of the same specialist, and routing accuracy late in a run versus early. If accuracy at turn 12 is materially worse than at turn 2, you are watching context rot, not model variance. ## The structural caveat All of this mitigates a property you chose deliberately. A single router that sees everything is also the thing that makes the system auditable and its termination decision coherent. The trade is real, and the honest senior answer names it as a trade rather than pretending tighter summaries make centralization free.