Orchestrator Patterns, Codex Speed, Kitesurf Browser — AI Daily Aug 06

325 messages · 61 active members

325

messages

active members

@jarvisballer, @sav310, @arielletolome

top contributors

Overview

Today's discussion centered on orchestration architecture, with @jarvisballer and @mb29266 sharing detailed playbooks for keeping orchestrators 'dumb' — offloading real work to fragmented subagents governed by rigid SOP markdown files, 1-hour max run caps, and handoff docs written before 20% context compaction. Consensus: orchestrators that try to be smart cause scope creep, infinite loops, and burned tokens. Stripping write/edit tools helps but isn't foolproof since agents route around via bash. Speed complaints about Claude Opus 5 pushed several builders back to Codex, which many reported as 2-4x faster (possibly due to Cerebras integration). GPT 5.5 medium emerged as a favored 'dumb' orchestrator, while Fable, Grok 4.20, and Opus 4.8 remain top copywriting picks. Multiple members flagged Opus feeling degraded overall, with /fast mode suggested as a workaround. On the tooling front, Cloudflare quietly launched Kitesurf — a Rust-based agent-first browser running in V8 isolates on Workers — alongside cloudflare-os for agent workspaces. Builders also swapped A/B testing stacks (Posthog, Funnelish, custom Astro + Vercel rewrites with cookie-stamped variants), shared open-source AI job search agents (career-ops, AIHawk), and passed around the leaked Claude Design system prompt and Impeccable for frontend polish.

Topics

Orchestrator + Subagent Architecture Patterns45 msgs

@jarvisballer and @mb29266 detailed a pattern where the orchestrator reads a rigid SOP markdown, delegates everything to fragmented subagents (writing, editing, formatting kept separate to avoid context poisoning), and enforces 1-hour max runs per coder. Handoff docs at 20% pre-compaction and subagents in separate terminals allow reattachment to fresh orchestrators. Stripping write/edit tools helps but agents still work around via bash.

Codex vs Claude Opus Speed & Quality24 msgs

Multiple builders reported Opus 5 running dog slow with degraded output quality, while Codex feels 2-4x faster than a month ago, possibly linked to Cerebras integration. @jasonakatiff and @scalingfrog swapped back to Codex; @KtargetMedia noted Claude 4.6 is currently faster than 5. /fast mode was suggested as an Opus workaround, and fresh sessions help all models stay usable.

Cloudflare Kitesurf & Agent-First Browsers10 msgs

@drcopybymatt flagged Cloudflare's Kitesurf launch — an agent-first browser built in Rust (not Chromium) running in V8 isolates on Workers. @sav310 shared cloudflare-os, an agent workspace for building 'gadget' apps with centralized reporting. A member also shared their own Rust headless screenshot browser project that recently pivoted back to Chromium for stability.

A/B Testing Landing Pages: Tools and Custom Stacks5 msgs

Members compared approaches to split-testing landing pages, with recommendations for Posthog, Funnelish, and Checkout Champ's native functionality. One member shared a detailed hand-rolled Astro implementation using Vercel serverless functions, cookie-stamped variants, and Astro.rewrite to keep a single paid-traffic URL while serving different arms.

Copywriting Models & PRD-Driven Long Runs30 msgs

@raresmol prefers Fable for long-form but rotates GPT, Grok, and Opus when Fable starts lecturing; @fuckyesiwannatalkbusiness pushed Grok 4.20 in unhinged mode. @jarvisballer spends 2-3 hours upfront on PRDs with full interactive HTML mockups, then leaves builds unattended 16+ hours to reach ~90% completion. Cross-model adversarial review (Codex builds, Opus reviews and vice versa) plus 1-hour run caps prevents drift.

Key Takeaways

Keep orchestrators 'dumb' — delegate everything to subagents via rigid SOP markdown and check state after each call
Cap subagent runs at 1 hour and write handoff docs before 20% context compaction to survive orchestrator restarts
Codex is currently 2-4x faster than Claude Opus 5 for many builders; try /fast mode if you must stay on Opus
Fragment skills by function (writing / editing / formatting stay separate) to prevent context poisoning
Cross-model adversarial review (Codex builds → Opus reviews, or vice versa) beats same-model self-review

Hot Threads

@jarvisballerstarted

Orchestrator + subagent architecture with 1-hour run caps and handoff docs

40 replies8 participants

@mlsmdmstarted

Fable and Opus slowness — many switching back to Codex

20 replies10 participants

@Andrenator8started

How is everyone running A/B tests on landing pages?

5 replies4 participants

Linked Items

GitHub - santifer/career-ops: Open-source AI job search: scan job portals, evaluate listings with a structured A-F rubric into a 1.0-5.0 score, tailor your CV, track applications — runs locally in your AI coding CLI (Claude Code, Codex, OpenCode, Antigravity…) · GitHub

github.com

@basant

GitHub - feder-cr/Jobs_Applier_AI_Agent_AIHawk: AIHawk aims to easy job hunt process by automating the job application process. Utilizing artificial intelligence, it enables users to apply for multiple jobs in a tailored way. · GitHub