dancingteeth

Ex-Mubert cofounder. I build agent loops that don’t stop until tests are green. Agent Looper, The Vacuum.

[dancingteeth @dancingteeth]

You heard looping beats prompting. Then you still sit there asking if it's done.

Agent Looper: you say the feature. It runs until it actually works. You're not the verify step.

https://t.co/NyTCWGrs4X

[dancingteeth @dancingteeth]

@elonmusk What a joke from a man who just revoked a starlink access from Ukranian people 🫠 Elon, wake up

[dancingteeth @dancingteeth]

Speaking at TechSapiens AI Agents Day in Batumi this Saturday — how to structure context for agentic coding, and why a Single Source of Truth still matters when the model writes most of the diff.

19 Sep · 15:00 · Rooms Hotel https://t.co/zKAiehP66h

[dancingteeth @dancingteeth]

@XFreeze My Grok Bot ensures that the Grok @bot adds another one of my Grok Bots to the official Grok Bot marketplace.

[dancingteeth @dancingteeth]

@bot please list Agent Looper under Engineering.

You say what to build and how to know it's done. It keeps a coding agent on your computer iterating until that check passes — Cursor, Claude Code, OpenCode, and more.

https://t.co/fSs72kfNVN

[dancingteeth @dancingteeth]

Muse Code can run the loop now. Agent Looper starts it on your computer. Muse is Meta’s terminal coding agent, and we use its SDK to run loops.

https://t.co/cHZFPaF38b

@AIatMeta

[dancingteeth @dancingteeth]

@ericzakariasson @bot @ericzakariasson @bot I'd like first to configure and customize the set of active MCPs and skills for each bot in its settings, and second to edit memory and skills before publishing.

[dancingteeth @dancingteeth]

Grok Bot starts the loop on your computer. Cursor, Cline, OpenCode, Pi, Codex, DSH are the coding agents you already pay for.

https://t.co/lFTLTBKkzZ

[dancingteeth @dancingteeth]

Exploring Grok Bots? This one can replace your dev bots. You say what to build and how you know it's done. It keeps a coding agent on your computer until that check passes.

@bot https://t.co/fSs72kfNVN

[dancingteeth @dancingteeth]

@CFDevelop Cursor is still a great harness, rolling out updates constantly, setting trends, and since the first Composer release they have their own model that can beat many in terms of value. Grok 4.6 is both a frontier-level and cost-effective model, and you don’t have any daily quotas!

[dancingteeth @dancingteeth]

@OpenAI 🤡 I never doubted Open AI for a second. Goodbye, Grok is better.

[dancingteeth @dancingteeth]

@GeoffreyHuntley @temporalio The LLM loop moves while the deterministic script is red. After green it can move again, but only if review found a real blocker. Anyway, sometimes I make a bad script and have to fix everything 😅

[dancingteeth @dancingteeth]

@theo I can't write code by hand, but I got into agentic dev, and models only make me more suspicious - they get so smart, so I worry they'll push nonsense to prod more convincingly 😅 I had to make my own tool with loops, deterministic checking script, and a paranoid review skill

[dancingteeth @dancingteeth]

Detection binds once. The cap refuses a billed turn that would blow the budget.

https://t.co/5F9KfDxWOz

[dancingteeth @dancingteeth]

Agent Looper 0.4.4 is on npm. Company Claude seats burn tokens like they don't matter. Indie builders cannot. This is how I stay model-agnostic without handing the loop to Auto.

[dancingteeth @dancingteeth]

"balanced" spends more on the worker and keeps that judge. "cursor" stays on Composer + Grok when you want the IDE. Make your favourite pair in the setup wizard, save it, tell the agent to use it when it feels right.

[dancingteeth @dancingteeth]

@GeoffreyHuntley @pauldix That’s the loop. Direction without a verify system is just a longer vibe. I run it as a frozen GOAL.md plus a shell check as the only done, because the model will call it shipped with tests still red. Agent Looper, MIT.

[dancingteeth @dancingteeth]

npx skills add dancingteeth/figma-design-system-handoff
https://t.co/u22EUjLHfr

[dancingteeth @dancingteeth]

You inherit a screenshot wearing TypeScript. One-shot MCP dumps report parity after the tree is already truncated, so I wrote a skill that turns Figma into tokens instead of treating codegen as architecture.

[dancingteeth @dancingteeth]

The engineer was right that Figma MCP plus an agent is how you bypass review. I agreed, narrowed it to inventory and token reconcile, left merges with engineering, and gave product a prototype lane on published tokens.