fx.sh adds multi-model subagent orchestration
Today's top 11 insights for PM Builders, ranked by relevance from X, YouTube, and LinkedIn.
fx.sh adds multi-model subagent orchestration
#1 đť•Ź
Guillermo Rauch announced that fx.sh can orchestrate subagents with different models and reasoning efforts—for example, Fable planning and Grok executing. Preferences can be set in AGENTS.md or a prompt, with users able to steer and interrupt agents; according to Rauch, it works with any model or gateway without server-side routing.
#2 đť•Ź
Qwen announced that Qwen3.8-27B is now running on Cerebras, with rapid inference available to try.
#3 đť•Ź
Dario Amodei announced his new essay, “We Must Pace the Frontier,” arguing that the AI industry should slow down through a three-part plan. Anthropic is unilaterally committing to the first step by giving third-party evaluators permanent, employee-level system access to verify safety adherence, report incidents, and assess model alignment during training.
Also covered by: @Sam Altman, @Demis Hassabis
#4 đť•Ź
Harrison Chase shared LangChain’s “How to Build a Custom Agent Harness,” describing it as a great blog by Sydney Runkle. The post quoted Garry Tan on systems of record versus domain-specific harnesses.
#5 ▶️
I Rebuilt Viral Ads With AI - Here's The Full Workflow (VEED Tutorial)
Helena Liu
Three viral ad formats were rebuilt with Claude, VEED, Kling, Seedance, and Veo: a paper-reveal ad, a confrontation-to-burger-shop ad, and an AirTag-in-a-chicken realtor ad assembled from generated clips.
- Claude watched the first roughly 6-second paper-reveal clip and wrote a video-generation prompt; in VEED’s Generate Video tab, the portrait orientation was selected, Kling was limited to 5 seconds, Seedance could generate up to 15 seconds, and one generation cost 40 credits.
- After about 30 minutes of prompt and model testing, Seedance produced the preferred paper-ripping reveal; Kling was reported as unsuitable for generated text, while uploading a reference image made the generated video more accurate.
- VEED was connected inside Claude through the MCP URL
veed.io/api/v1/mcp; Claude listed VEED avatars and more than 500 voices, generated talking-head clips, and stitched the chicken drumstick, animated tracking, and realtor-avatar clips into one ad in about 10 minutes.
#6 đť•Ź
There's An AI For That recaps Netflix’s VOID, an AI model that removes objects from video and corrects the scene’s physics afterward—a capability that distinguishes it from other object-removal tools.
#7 in
Guillermo Rauch shared that Tailscale’s model router uses Vercel AI Gateway as its underlying infrastructure.
#8 đť•Ź
Kevin Yien recapped using Muse to sequence three errands within one hour and place a Tarka takeout order ahead of time, including sending an order summary and payment-approval link. He described the experience as useful and frictionless, noting that the order was correct and he returned to the dance location with 3 min to spare.
Also covered by: @Kevin Yien
#9 đť•Ź
DeepLearning.AI recapped Andrew Ng’s latest letter in The Batch, arguing that top AI engineers shape what gets built rather than merely coding to spec. He highlights four core skills: driving rapid build-and-feedback loops, making product decisions, communicating across functions, and exercising high-agency ownership.
#10 in
Peter Yang shared a prompt that he said Grok Bot used to help him recover approximately $500 in unclaimed money by searching official state databases using his full name and past cities. The prompt requests each result’s type, amount, and who owes it, and Yang said it could also be tried with other agents.
#11 đť•Ź
Alexandr Wang said users can often ask Muse to connect to something, and it will write the entire integration in the background.