holy f*ck, SpaceXSI gave 5 accounting jobs to Grok bots
the team says it would need more than twice the headcount to do the same work without them
the numbers:
invoices: roughly 24 hours down to minutes
payment reconciliation: 3–4 hours down to 30 minutes
billing support: an estimated 8 team-hours freed up every day
five jobs already running:
account setup, invoicing, billing support, payment matching, customer inquiries
here’s the setup worth copying:
> give each bot one clearly defined job
> write its operating manual in Notion
> let it watch a person do the work
> make it propose actions before it can execute them
> let it handle routine work once it’s ready
their usual onboarding takes 3 weeks
every night, the bots reread their manuals. they cannot change the rules themselves
the customer-facing bot took 3 months to train. credits and refunds still need human approval
if you want to build agents for businesses, find the person spending 3 hours a day matching payments
ask them to show you exactly how they do it
ONE CLAUDE AGENT COST $9. THREE CLAUDE AGENTS COST $200. ANTHROPIC CHOSE THREE.
Prithvi Rajasekaran at Anthropic Labs published how they got Claude to build full-stack apps across multi-hour sessions.
The earlier harness used an initializer, one feature per session, and context resets. Sonnet 4.5 would wrap up early. Anthropic calls this "context anxiety." Opus 4.5 mostly stopped doing that, so this run drops the resets and keeps one continuous session, with the Claude Agent SDK compacting context as it grows.
Three agents, one job each:
> Planner takes a 1-4 sentence prompt and writes a high-level product spec
> Generator builds one feature per sprint on React, Vite, FastAPI, and SQLite, then commits to git
> Evaluator opens the live app with Playwright MCP, clicks through it like a user, and grades UI, API, and database state
Before any code, the generator and evaluator agree a sprint contract: what "done" looks like. They pass that through files. If any score falls under the threshold, the sprint fails and the generator gets a written critique.
Anthropic compared a solo agent with the full harness on this prompt: "Create a 2D retro game maker with a level editor, sprite editor, entity behaviors, and a playable test mode."
Solo agent: 20 min, $9
Full harness: 6 hours, $200
The harness cost over 20x more. Anthropic wrote that the quality gap was immediately apparent.
Save this.
Pick reasoning effort by depth (from OpenAI's Codex guide):
> Medium -> a quick survey
> High -> tracing one cause across several systems
> Extra High -> several causes that may interact
Sources:
https://t.co/8IGpOgfC0D
https://t.co/ME2Kme9kYK
Save this.
GPT-6.2 ASTRA STOPS AT "NEXT STEP" AND WAITS. YOU HAVE TO TELL IT TO FINISH.
Don't tell GPT-6.2 Astra "be concise."
Don't tell GPT-6.2 Astra "make it look nice."
Don't tell GPT-6.2 Astra "add tests for everything."
OpenAI's own GPT-6.2 guide says Astra works differently from older models.
Here are 6 prompts to copy and paste:
6. The Real-Size Check
"Match this reference image. Render it, check it at the size people will actually see it, fix what gets lost, then try every action it should support."
No reference? Ask Codex to make one with GPT Image first.