Hermes Stacks, Claude Outage & Codex Loops — AI Daily May 16

426 messages · 65 active members

426
messages
65
active members
@jasonakatiff, @arielletolome, @rmktg
top contributors

Overview

Saturday was dominated by @rmktg's 40-minute Hermes walkthrough video, which triggered a wave of builders rebuilding their agent stacks around multi-profile orchestrators (Frenchie, Craig, Wordsmith, Eddie, Lloyd, Chad), per-profile MEMORY.md compaction, and CLI-based handoffs replacing outbox polling. @Wootbro fed the video into Claude's new frame-extraction tool to auto-generate a full architectural upgrade brief. Parallel drama: @jonmacofficial discovered Higgsfield's 'Supercomputer' is allegedly 50 skills lifted straight from NousResearch's Hermes Agent — he's reverse-engineering the binary and already pulling preset prompts via CLI. Claude Code went down mid-Saturday — unusual for a weekend — fueling speculation about an imminent model release and renewed frustration with Anthropic's rate-limit tightening (@nowwatchthisdrive claims Max $200 effective inference value has collapsed from ~$5k to ~$200). Once it returned, builders compared Codex thinking-mode settings: @arielletolome runs 5.5 with extra-high thinking, @c_1media plans with Opus 4.7/xhigh then implements with Codex low, and @jasonakatiff detailed a 7-step Codex+Claude mutual-review loop that landed a major auth refactor. Gemini 3.2 Flash and Nova Pro emerged as serious fallbacks. On the practical side: @expadz published a detailed video-gen price breakdown (Kie.ai wins at $0.77/10s 480p), email infra converged on Resend for small senders and SES+Sendy for high-volume operators, @jasonakatiff's LeadRouter already closed a $4k IUL commission on day-one leads, and side threads covered n8n on Cloud Run, FB/Reddit scraping pipelines, and Qwen3.5-27B-DFlash signaling the local-model arms race continues.

Topics

Hermes Agent Stacks + Higgsfield Skill Heist40 msgs

@rmktg's 40-minute walkthrough of his Hermes setup (Frenchie orchestrator, Craig dev, Wordsmith copy, Eddie video, Lloyd admin, Chad media buyer) reset everyone's architecture — emphasizing capability-first builds, ~2200-char MEMORY.md caps with nightly compaction, and CLI handoffs over outbox polling. @Wootbro used Claude's video-frame tool to auto-generate an upgrade brief. Meanwhile @jonmacofficial caught Higgsfield's 'Supercomputer' shipping 50 skills allegedly stolen from NousResearch's Hermes Agent and is reverse-engineering the binary.

Claude Code Outage + Codex Reviewer Loops & Thinking Modes33 msgs

Claude Code went down Saturday, sparking model-release speculation and renewed anger about Max $200's allegedly collapsed inference value. @jasonakatiff detailed a 7-step Codex+Claude mutual-review process (parallel agents, gap analysis, Opus aggregation) for a successful auth refactor. @c_1media recommends Opus 4.7/xhigh for planning then Codex low for implementation; Gemini 3.2 Flash and Nova Pro (Sonnet-quality at ~25% cost) are emerging as fallbacks.

Video Ad Pricing: Kie.ai vs Higgsfield vs Poyo12 msgs

@expadz published a detailed cost comparison for Seedance/Kling: Kie.ai at $0.77/10s (480p) and $1.65/10s (720p), Higgsfield $0.65/$1.52, Poyo cheapest at $0.10/8s but unstable. Verdict: Kie.ai wins for non-subscription flexibility, 480p is fine for Meta/social, mix platforms to cherry-pick rates.

Email Infrastructure: Resend, SES+Sendy, Klaviyo Alternatives20 msgs

@samb69 asked for Klaviyo alternatives with API/MCP for programmatic campaigns. Community converged on Resend for small senders and Amazon SES + Sendy ($69 one-time) for high-volume operators. @robinroy is eyeing SES to recover 50k+ wrongly-unsubscribed Swiss customers Klaviyo won't let him re-import. Mumara and Cloudflare's new email service also flagged.

Scraping, n8n Deployment & Local Models14 msgs

@yangthegoat asked about autopiloting FB/Reddit comment scraping pipelines — suggestions included Cloudflare D1 SQL for queryable storage, Apify, and @mb29266 advocating self-built free scrapers. @cris_vels got n8n-on-GCP deployment options from @kennetele (Docker Compose VM to Cloud Run to k8s). @watchmedropship shared Qwen3.5-27B-DFlash as the new local-model trend.

Key Takeaways

  • Cap each Hermes agent profile's MEMORY.md at ~2200 chars with a nightly compaction cron — @rmktg's media buyer hit 91 skills before forced compaction was needed.
  • Kill outbox-polling crons for inter-agent communication and standardize on Hermes CLI's send-message verb — direct CLI wake-ups dramatically improved handoff reliability.
  • Codex split workflow: Opus 4.7/xhigh for planning, Codex low for implementation — but only when the plan is airtight. Use Codex+Claude as mutual reviewers, not solo builders.
  • Resend is the sweet spot for small-to-mid email senders; only move to SES+Sendy when volume justifies the operational pain. Nova Pro hits Sonnet quality at ~25% of the cost.
  • Kie.ai 480p at ~$0.77/10s beats Higgsfield's subscription model for Meta/social ad flexibility — 480p is good enough and not subscription-locked.

Hot Threads

@rmktgstarted

Hermes setup walkthrough video + orchestrator upgrade briefs

30 replies10 participants
@samb69started

Klaviyo alternatives with API/MCP for programmatic email campaigns

18 replies7 participants
@expadzstarted

Video generation pricing: Kie.ai vs Higgsfield vs Poyo

10 replies4 participants

Linked Items

ProtonMail's Urgent Warning: Never Sign In With Google

x.com
@arielletolome

Massive Facebook Ad Swipe File Leaked - AI, CGI, VSLs & More

x.com
@jonmacofficial

Track Every AWS SES Email: Open-Source Dashboard for Free

sessy.do
@joshozjoshoz

Gemini 3.2 Flash: 92% of GPT 5.5 Performance at 20x Lower Cost

x.com
@mattkowald

Unable to analyze - insufficient content data

youtu.be
@rmktg