OMP Harness, Grok 4.5, Skills & Sub-Agents — AI Daily Jul 15

519 messages · 74 active members

519

messages

active members

@samb69, @watchmedropship, @Kieran

top contributors

Overview

The community spent the day rearchitecting coding stacks around a new multi-model reality. OMP.sh emerged as the harness of choice, with @jcartu, @samb69, and Kieran championing it as a unified replacement for Claude Code and Codex — one interface across all OAuth and API accounts, per-role model auto-switching via agents.md, and routing even to local GPUs. A/B tests against CC favored OMP on terminal UX and sub-agent orchestration. Model loyalty is shifting fast. Fable 5's aggressive guardrails are burning tokens before refusing routine dev work, pushing builders toward GPT-5.6 Sol for planning/review and Grok 4.5 as the fast executor — @samb69 calls Grok 4.5 'Opus 4.7 level' minus the political correctness. Sonnet 4.5 still wins for copy, and Fable on low reasoning effort is emerging as a token-efficient sweet spot. Grok Build was also open sourced today, and Claude, Codex, and Grok all reset weekly limits simultaneously, sending builders back to their keyboards. @geilt made a detailed case for building skills and specialist sub-agents over relying on session history — a handbook at your desk beats re-figuring things out daily. Infrastructure rounded out the day: @jcartu's local GLM 5.2 on RTX 6000s hits ~110 tps and pencils out for teams of 6–8, Hetzner auction boxes host Hermes + OpenClaw + CC agents that skill-share over Telegram, and @watchmedropship demoed a Hermes-driven canvas ad builder alongside agent-orchestrated video pipelines using Grok Imagine at ~12¢/clip.

Topics

OMP.sh Becomes the Default Multi-Model Coding Harness22 msgs

Members A/B tested omp.sh against Claude Code and reported it wins on terminal UX, sub-agent deployment, and unified access to Gemini, Grok, Fable/Opus, and GPT under one roof. Kieran uses an agents.md to route tasks per model (e.g. GPT-5.5 for code review), and @jcartu confirmed it can also route to local GPUs. Most are dropping CC in favor of it.

Fable Guardrails Drive Rotation to Sol, Grok 4.5, and Sonnet42 msgs

Fable 5's broad safeguards are flagging routine dev and sysadmin tasks, burning tokens before refusing. Builders are rotating to GPT-5.6 Sol for planning/review and Grok 4.5 as the fast executor, while Sonnet 4.5 still wins for copy. Fable on low reasoning effort is emerging as a token-efficient favorite when it doesn't refuse.

Building Skills and Specialist Sub-Agents18 msgs

@geilt laid out a compelling case for creating skills and specialist agents rather than relying on session history — comparing it to having a handbook at your desk vs re-figuring things out daily. Members shared examples like an ADB expert agent for Android telephony testing, and discussed using Fable to spawn skills that lesser agents then execute in parallel panes with iterative improvement loops.

Self-Hosted GLM 5.2 and Hetzner Agent Stacks34 msgs

@jcartu detailed a local GLM 5.2 setup hitting ~90–120 tps across 8 RTX 6000s — a bargain when serving 6–8 devs 24/7 with no guardrails and Opus-level quality. Meanwhile @sibunting and others run Hermes, OpenClaw, and multiple CC tmux sessions on Hetzner boxes (auction servers at hetzner.com/sb flagged as best value), with agents skill-sharing over Telegram. Hardening on day one is non-negotiable.

Grok CLI Open Sourced + Weekly Limit Resets22 msgs

Grok Build was released as open source just as Claude, Codex, and Grok all reset weekly limits on the same day. @c_1media shared a setup using Codex 5.6 high as a main planning thread spawning sub-sessions running Grok CLI in separate worktrees. @Tz1888 reported running six long sessions after resetting at 5%, and builders joked about the 'golden age' of compute before providers tighten the leash.

Key Takeaways

OMP.sh is winning as a unified harness — one CLI across Claude, Grok, GPT, Gemini, and local GPUs with per-role routing via agents.md.
Emerging meta: GPT-5.6 Sol or Fable for planning/review, Grok 4.5 as the fast executor, Sonnet 4.5 for copy — Fable's guardrails are actively costing builders tokens.
If you catch your agent repeatedly searching for the same info, that's the signal to build a skill or specialist agent — pre-prepared reference docs beat stochastic re-derivation.
Local GLM 5.2 on RTX 6000s only pencils out at team scale (6–8 devs); Hetzner auction servers beat standard pricing for dedicated agent hosting.
Grok CLI + Codex orchestration in separate worktrees is emerging as a viable multi-agent pattern, especially with weekly limits freshly reset.