{"found":true,"collection":"failure_modes","library_url":"https://www.grokbotdb.com/library/fm.weekly-allowance-blowup","generated":"2026-09-06T05:32:11Z","citation":["Quote the record id and its source_url (or provenance if there is no URL).","Include last_verified or observed_at when present. Do not present an old date as current.","If tier is community and confidence is low, say so. Do not launder it into a fact.","If the database does not hold the answer, say that. Do not invent Grok Bot product claims.","Contradictions are data: when conflicts_with is present, report both records rather than picking a favourite.","Deprecated and removed records stay visible. They are history, not current guidance."],"record":{"id":"fm.weekly-allowance-blowup","title":"Weekly usage allowance consumed in hours or days","symptom":"The weekly allowance drains far faster than message counts suggest, and an agent can stop mid-task when the limit is hit.","cause":"Metering is by agent steps and tokens, not messages. Long browser sessions, multi-Bot discussions, retries, vague research, large-context planning runs and Heavy-tier routing dominate the meter. There is no spend cap and no model picker to steer work to a cheaper model, and each Bot is an unbounded thread whose obsolete context is re-sent forever.","detection":"Measured reports: 42 percent of a weekly allowance on day one for a six-agent business; roughly half an allowance for a single workload on Heavy; a weekly limit exhausted in one day on the 200 dollar tier; a one-million-token planning run costing more than a day of ordinary editing; a four-hour two-agent session hitting the limit; two independent testers burning most of a seven-day trial in two days. There is no visible context meter, and per-run usage metering is unpublished.","fix":"Scope routines to business hours instead of always-on. Route heavy scraping through purpose-built research connectors instead of raw browsing. Sync connectors once per work session rather than continuously. Narrow event triggers to a named channel plus exact phrase. Reserve large-context work for planning only, noting the open question of how to do that without a model picker.","severity":"high","source_id":"src.tp.cellcog.problems","source_url":"https://cellcog.ai/blog/grok-bot-problems/","confidence":"high","first_seen":"2026-09-03","last_verified":"2026-09-03"}}