Copy this Grok Bot configuration: four seats, half the cost
Ling publicly released the Grok Bot Harness configuration, using a folder as a manager and four seats to choose models by cost; jev handles routing and completion decisions. Builders never self-assess; verifiers write receipts from scratch. A five-round run costs $0.82, half the cost of giving everything to grok 4.7. Full configuration logic included.
Lauren from SpaceXAI has open-sourced her entire agent stack: 102 skills that can be added to Grok Bot in seconds. Her team ran it 10,000 times in the first week. Many people give a bot a prompt and hope for the best; she gives every task a playbook, every change a reviewer, and every decision a rule. These skills are grouped into categories such as Getting Started, Understanding, Building, Review, Writing, and Agents, plus 24 principles and 23 playbooks.
Even more worth copying is the Grok Bot Harness configuration that Ling released. She uses a folder structure as a “manager,” solving the problem that Grok Bot has no manager type. Four seats, each choosing the most cost-appropriate model, with files as memory that persists after every conversation.
The folder is at /workspace/grok-harness. Every bot reads from and writes to it, handing off through it. Structure:
- GROK.md: the contract every bot must follow
- advisor.md, scout.md, builder.md, verifier.md: the four seat files
- jev_seat.py: asks jev which seat should get the next step
- jev_done.py: asks jev whether GOAL.md has been achieved
- GOAL.md: the finish line, written as checks a stranger could verify
- receipts/: a proof file for every passed check
The four seats map to four models:
- Advisor: grok 4.7, owns GOAL.md and the plan, never writes code
- Scout: grok 4.3, long context, reads the codebase, never modifies it
- Builder: grok build 0.1, writes code from only one task card at a time
- Verifier: grok 4.7, fresh context, only sees GOAL.md, the diff, and the task card, never the build conversation
Why split it this way? Reading code mainly consumes input, so the scout uses a cheap long-context model. Writing code mainly consumes output, and the model used by the builder costs only one third of grok 4.7's output price. The verifier uses the strongest model and starts from zero; this is intentional. A checker who watches the work being done is essentially grading its own homework.
Before each card starts work, it first goes through jev_seat.py. jev returns only three things: the choice, the confidence, and the probability of each option. If confidence ≥ 0.80, the card goes to the corresponding seat; if below 0.80, it goes to the advisor to decide. In one run, 14 routing calls cost about one tenth of a cent. For the same calls, having grok 4.7 write them as text would cost about 8 cents.
At the end of each round, jev_done.py judges whether GOAL.md has been met. If done ≥ 0.92, stop; if below 0.92, the advisor takes the weakest check and moves to the next round. If it still hasn't stopped by round 6, everything pauses and goes to a human.
The receipt system is the hardcore part of this configuration. After each check passes, the verifier writes a receipt file containing the command, output, commit, and time. Builders never write receipts. jev_done reads the receipt list; saying “done” in chat doesn't count—no receipt means not done.
Cost comparison: one run for one feature, five rounds, costs $0.82 with seat splitting, versus $1.58 if everything is handed to grok 4.7. The difference is mainly in the builder's output: 66,000 tokens are billed at $2 per million; with grok 4.7, that would be $6. Scouting is kept under 200,000 tokens each time, because at 200K the entire request is billed at double.
Build order: folder, GROK.md, GOAL.md, four seats, jev_seat, jev_done, and finally the group chat. The group chat holds six bots: advisor, scout, verifier, builder-1, builder-2, builder-3. Each bot's screen can run only one computer-operation task at a time, so three builder bots are needed in parallel; one bot opening three tabs is useless. Bots hand off through the folder and don't need to move conversations.
Ling also provided 11 prompts, from creating the folder to the Friday audit, ready to copy directly. The prompt for starting a run requires one line per round in the group chat: round, passed checks / total, P(done). The Friday audit tallies the number of runs, the average P(done) at stop time, cards routed to the advisor, checks that failed twice, and any bot that worked outside its seat.
The counterintuitive point of this configuration is: use the cheapest models for lots of reading and writing, and reserve the most expensive model for independent verification. The verifier never sees the builder's conversation and only trusts receipts. Beyond saving cost, it gives quality verifiable evidence. If you're also running Grok Bot, try copying this folder tonight and writing your first GOAL.md tomorrow.
发布时间: 2026-11-22 20:35