Grok as Main Coder, OMP Harness, Multi-Agent Stacks — AI Daily Aug 28

268 messages · 63 active members

268
messages
63
active members
@jasonakatiff, @samtome, @samarh90
top contributors

Overview

Wednesday's chatter centered on model selection and harness maturity. A vocal thread pushed Grok as the main coder — fast, clean, and less preachy than Claude — with Jason Katiff, Jon Mac, and others running Grok Build CLI alongside Claude Code or Codex for review. Others held firm on Claude Max as the safe bet, while @leewardbound revealed he's been running Terra as his OMP default for months, arguing tighter upfront planning with a lighter model beats vague grand prompts to Sol. OMP + Orca continues to dominate the workflow discourse. @jonmacofficial published a Supercut walkthrough of his Orca/OMP flow with tight RULES.md constraints (240-char chat cap, no preamble, spawn specialists via task), and @jasonakatiff dropped a Lead Router office hours video showing layered iteration through dev → branch preview → staging → prod pipelines. Discussion also surfaced Darkbloom (rent-out Mac inference) and PhoneLLM, an open-source voice agent fine-tune of Nemotron Nano 30B. The deeper thread was @thewildzeno's progression map — prompt → context → harness → loops → graph — framing graph engineering as the current endgame for software factories, where agents are right-sized on intelligence, cost, and speed at each node. Jason Katiff agreed, calling his version 'workflows vs graphs.' Side threads covered Veo vs Seedance vs Omni for UGC ads, Meta Managed Accounts as a fix for banned personal profiles, and bypassing AI text detectors by rewriting at 6th–8th grade level.

Topics

@jonmacofficial and others are running Grok (xHigh reasoning) as their primary coder, calling it fast, clean, and less prone to lecturing than Claude — great for 1:1 funnel clones and native-style hooks. Kieran and others stayed on Claude Max as the safer long-term bet, while @jasonakatiff runs a mixed fleet: 5 Claude Max, 1 Codex Max, 1 Grok Super Heavy. Consensus: multi-agent review (Grok codes, Sol/Codex reviews) beats letting one model rip.

@jonmacofficial shared his full Orca/OMP flow with RULES.md constraints — 240-char default chat cap, no preamble, decisions routed through the ask tool, and mandatory ponytail-auditor on plans/PRs. @leewardbound is running Terra as his OMP default, arguing lighter models plus better upfront instruction beats heavy planning with Sol. /advisor in OMP got flagged as a killer feature.

@thewildzeno mapped the AI engineering progression: prompt → context → harness → loops → graph. Graph engineering means isolated agents in defined flows with governance and boundary contracts, letting you right-size intelligence, speed, and cost per node. He's building a software factory now; Jason Katiff calls his version 'workflows,' and both agree it's overkill for vibe coding but the definitive path for long-term systems.

Kieran's team is switching from Veo (which has 'been printing winning ads since late last year') to Omni for UGC — 8-second Veo clips stitched together with careful scripting have kept them above the AI slop line. @startropics reported Seedance won for their joint pain offer using 3-sec hooks mashed with real footage. Face detection bypasses via the Reddit sketch method also got shared.

@scalingfrog is using Grok to review GPT Sol output by comparing goal to actual result — Jason's earlier point that multi-agent review beats single-model runs is landing. @jrizzolo is testing Luna Max for basic testing and Sol for bigger issues to avoid endless review loops. @jaypozo dropped UltraCode entirely for Sol-vs-Sol, and Fable-plans/Grok-codes/Sol-reviews emerged as a popular triangle.

Key Takeaways

  • Grok on xHigh reasoning is emerging as a legit main coder — fast, clean, and less preachy than Claude, especially for funnel clones and native hooks.
  • Multi-agent code review (Grok codes → Sol/Codex reviews) beats single-model runs; reviewer loops can be tamed by right-sizing the reviewer model.
  • OMP's edge: model-agnostic mid-session swaps with LSPs fixing code in the background; enforce 240-char chat caps and RULES.md to avoid essay outputs.
  • The engineering progression is prompt → context → harness → loops → graph — graph engineering is the endgame for cost/intelligence optimization per node.
  • Meta Managed Accounts (via work.meta.com) offer business-only personal profiles as a potential fix for banned ad accounts and farming issues.

Hot Threads

@jonmacofficialstarted

Grok as main coder + OMP RULES.md workflow

18 replies9 participants
@thewildzenostarted

Graph engineering as the AI dev endgame

11 replies5 participants
@leewardboundstarted

Terra as OMP default beats Sol grand-planning

14 replies4 participants

Linked Items