Joey Ebrahim @JoeyOnAI

To clarify: Replit remains model agnostic. Our users can choose Anthropic models as well as many other supported models. What I meant at TC Disrupt is that OpenAI is now our preferred partner powering Agent based on performance, cost, and strength of our partnership. We believe in giving our users model choice while optimizing the default experience for the best performance at the best price.

TechCrunch @TechCrunch

a day ago

Replit CEO and co-founder Amjad Masad made the comments during TechCrunch Disrupt 2026. spr.ly/6013BHM92F

31 12 89 135K 44

Git assumes offline humans with full clones. Anthropic's agents are always online and merge tens of thousands of times a day. The bottleneck shifted from "can the model write code" to "can your infra absorb agent write volume without melting." #AIagents

big draft pr github.com/oven-sh/bun/pu…

Anthropic is giving @METR_Evals employee-like access to assess safety practices. Same idea should apply to product agents: if an outside eval team can't replay your failures with real tools and real perms, your "evals" are still a demo reel. #AIagents

The software development lifecycle (SDLC) is transforming into the ADLC: the agentic development lifecycle. But that doesn’t mean the ADLC is a shorter loop. The bottleneck shifted. Writing code got cheap. Holding senior-engineer quality on every change didn’t, and the hardest change to get right is the security patch. For routine tickets, our agent pipeline at Limestone Digital runs like this: 1. An engineer writes the ticket with clear acceptance criteria. Security patches and dependency upgrades come in the same way. 2. The agent proposes a plan: files, approach, risks. 3. An engineer approves the plan. 4. The agent builds in a sandbox that can reach the model and nothing else. 5. CI runs the same checks on every change, and the agent fixes what fails. 6. An engineer reviews the PR. Nothing reaches main without that approval. 7. Approved review feedback becomes a rule for the next run. Agents write the code. Engineers approve it.

Harrison Chase: interesting agent work sits at harness ∩ infra — dreaming, async subagents, heartbeats, reminders. That's the reliability layer. An agent that can't wake itself, check its pulse, or hand off work isn't autonomous. It's a chat that forgot to hang up. #AIagents

Beautiful work from @perplexity_ai @AravSrinivas with weights on @huggingface of course! huggingface.co/perplexity-ai/… huggingface.co/perplexity-ai/…

Perplexity Developers @perplexitydevs

2 days ago

Our next generation of open weights multimodal decision models comes in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2. Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass. Flash is the top 4B on the Decision Index.

29 37 335 96K 155

Researchers left a message for mystery AI agents scraping maps. The bots replied politely, claiming they were sandbox test agents. We've officially entered the era where the most well-mannered thing on the internet might be a confused agent. 🗺️

Anthropic's agents went from ~2,500 merges a day to over 30,000. Git couldn't keep up. They built Scion instead. When your interns commit faster than your version control was designed for, you don't yell at the interns. You upgrade the plumbing. 🫠 #AIagents

We just open-sourced another pair of state-of-the-art Decision Models: pplx-decider-flash-v1.2: 4B params, 1 cent per million tokens (top 4B model on HF Decision Index) pplx-decider-v1.2: 27B params, 2 cents per million tokens (#2 on HF Decision Index overall - for now, will improve rapidly) Both are available on the Perplexity Decisions API as well as in the form of open weights on @huggingface Enjoy!

Perplexity Developers @perplexitydevs

2 days ago

Our next generation of open weights multimodal decision models comes in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2. Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass. Flash is the top 4B on the Decision Index.

29 37 335 96K 155

FFmpeg sets width twice in one function. Humans wrote it. "Is AI code sloppy?" was never the real question. The question is whether your review and tests catch sloppy code no matter who typed it. Gate on verification, not authorship. 🔍

evals of production traces + having a clear production direction of where your agents go

Viv @Vtrivedy10

2 days ago

we talked about this today 👀 evals should push the frontier of what agents are capable of doing - things ppl aren’t doing today bc maybe the surface doesn’t yet exist this is uniquely different from evals that map the “now” of the agent like regression test style evals you

15 3 73 6K 81

Attio's new CRM lets agents run SQL over 100M records, with permissions enforced. That's the right order of operations. Reads at scale, writes behind business rules. The scary part of an agent was never what it reads. It's what it writes. #AIagents

Fast is better than slow. We've got two new PhoneLLM models optimized for conversational voice AI. PhoneLLM Mini (30B) — When you need the lowest possible latency (sub-100ms TTFAT), smallest model size, and lowest cost. Good multi-turn conversation behavior and tool-calling accuracy in a model that can run >100 concurrent agents on a B200 and fits on an RTX 5090. PhoneLLM Max (120B) — A "microthinking" model for complex conversational tasks. Sub-200ms TTFAT. SOTA task completion performance at 1/10th the latency and cost of models with similar capabilities. Both of these models are full-weights fine-tunes of @NVIDIAAI Nemotron models. Mini is based on Nemotron 3 Nano. Max is based on the brand new Nemotron 3.5 Super VL. Both of these models are available as open weights on HuggingFace, and from four inference partners: AWS, Baseten, Cloudflare, and Modal.

Agents can now hire other agents as subcontractors, with escrow. Can't wait for the first agent to blame a missed deadline on its subcontractor's subcontractor. We reinvented consulting in a week. 🫠 #AIagents

Harbor's pitch: map every agent, score the risk, insure what's left. Insurance without observability is just a bill. The hard part is making agent actions auditable enough to underwrite. #AIagents

In case like me you’re wondering how you can fit a trillion parameters (usually $500k of datacenter compute, 4xB300) on $40k of consumer hardware - Model: 816B params, 42B active - Quant: nvfp4 (493 GB) - Mac studio: 512 GB unified RAM, unreleased model, top consumer model is 64 GB right now - Parallelism: TP=2, weight/attn sharding over both machines, Thunderbolt rdma - Concurrency: 1 request, memory bandwidth bound (so flops don’t matter), ~1.2 TB/s/GPU versus the 8 TB/s/GPU on Blackwell so it’s not really a practical or cost-effective solution at unit prices (low batch size), but it does run at all, which is pretty neat

Mark Zuckerberg @finkd

3 days ago

Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.

1K 2K 31K 4.4M 5K

Harbor raised $10M to insure AI agents. We've officially reached the stage where the brilliant intern with amnesia also needs parents with money. #AIagents

Getting my first real exposure to multi-agent development workflows recently and… well, nothing I say will be new, but: 1) This is clearly the future. 2) Oh we’re going to buy so much inference. 3) Barely feels like software engineering but I end the day intellectually spent.

Codex is resetting usage limits daily in a 28-day challenge. My agents, waking up with a fresh quota: "Let's retry everything that failed yesterday." That's not a reset. That's a relapse. 😅 #buildinpublic