---
name: "Token budget — kill amnesia, route models"
description: "Use when Grok Bot / agent spend is high, context is bloated, or you want more agents on the same token budget."
---

# Token budget — kill amnesia, route models

## When
Use when agent spend feels 10× too high, a single chat is chewing the budget, or you want many specialist bots on the spend that used to fund one fat agent. Trigger on “burning tokens,” “same spend / more agents,” or before spinning a fifth bot that redoes research.

## Stop
- Do not add more bots that repeat the same research under new names.
- Do not paste full chat history into every call “just in case.”
- Do not run grunt work on a frontier model when a cheap/fast model clears the bar.
- Do not invent a model picker if the host has none — use host fallbacks (e.g. Grok Build CLI for heavy jobs; keep the chat for decisions).
- Never post, pay, deploy, or send without the human’s explicit yes.

## OVEN
Default: Grok Bot tools — Shell / Grok Build (`grok -p`) for token-heavy compute / Task executors for parallel grunt / GenerateImage only when needed / connectors for live data. (Swap this one line for Astra / Fable / Claude.)

## Steps
1. **Name the waste.** List open bots and last 24h jobs. Flag duplicates (two bots researching the same topic) and any chat that keeps re-sending a giant transcript.
2. **Assume amnesia.** Treat every model call as if it forgot everything: pass only the minimal packet (goal, constraints, file paths, success check). Prefer files on disk over pasting history. Strip prior turns unless the decision depends on them.
3. **Plan once, then execute.** Write a short stepwise plan (or paper ticket) before the expensive loop. One plan → many cheap steps beats discovering the plan inside a frontier turn.
4. **Route by job hardness.**
   - Grunt (summarize, format, classify, scrape, draft from notes) → cheap/fast model or Grok Build headless.
   - Hard (architecture, ambiguous judgment, final copy lock) → frontier / main chat.
   - If the host has no model picker, route by *surface*: chat for decisions; `grok -p` / executor for bulk; never both on the same fat context.
5. **One bot, one job.** Split roles so specialists don’t re-pull the same context. A lead/CoS routes; specialists execute. Add a new bot only when a job already repeats without the human.
6. **Tighten prompts.** Precise scope, numbered steps, crisp done-when. Vague prompts multiply retries (and tokens) 10×+.
7. **Meter.** After one real job, note tokens or dollar spend. If still hot, cut context packet size again before changing models.

## Done when
- Duplicate research bots are merged or killed.
- Heavy loops run on the cheap path; decisions stay on the expensive path.
- Each specialist has one job and a minimal input packet.
- A short note exists: what was eating the budget + what changed.

## Cross-model
Recipe is fixed: amnesia-aware packets, plan-then-execute, route by hardness, one job per bot. Only the OVEN line swaps (which CLI / subagent / model names the host exposes). Grok Bot hosts without a model picker still follow steps 2–6 via surface routing.

## Source
Reworked from Mahax / SpaceXAI engineer token-budget breakdown discussed in https://x.com/Mahaximus_/status/2096234675794501686 (video + thread summaries). Not a transcript — recipe distilled for Grok Bot installs.