A permitting agent receives verbose GIS tool dumps that threaten the next reasoning turn. What context step fits?

Pruning verbose GIS tool dumps before the next reasoning turn is the correct context step for permitting agents because raw coordinate-heavy payloads consume tokens without proportional decision value. After a GIS tool returns parcel geometry, overlays, and attribute sprawl, the agent should retain the facts needed for the permit question—zoning codes, setbacks, floodplain flags—and discard verbose residue. Pruning protects Claude's remaining context for deliberation, citation, and resident-facing explanation instead of drowning the turn in coordinates.

Summarizing GIS parcel results into the system prompt for reuse fails because the system prompt is meant for standing instructions, not per-turn tool output; stuffing parcel summaries there bloats the prompt and leaves the verbose residue unpruned before Claude's next reasoning turn. Caching detailed GIS coordinate payloads in working memory fails conceptually—it retains the verbose residue in memory rather than stripping it out before the next reasoning turn, so the bloat and attention dilution remain. Expanding the model context window to retain GIS dumps fails because the problem is too much low-value GIS data already in context, not too little—making the window bigger just lets an oversized, unpruned payload fit rather than fixing the bloat.

Exam caveat: over-pruning can drop a critical overlay (for example, historic-district status); define a required attribute checklist per permit type before aggressive truncation. Operational check: capture a real GIS tool response, apply the pruner, verify required zoning and hazard fields survive, confirm token count drops substantially, and ensure the next Claude turn still reaches the same permit determination on the eval set.