Quick Hits

• A protocol for agents talking to businesses. @btaylor announced the Personal Agent Protocol, an open standard from Meta and Sierra with Genesys, instinct, RocketOTD, Shopify, Stripe, and Walmart as partners, defining how personal agents interact with businesses. @sriramk reads it as an OAuth-in-the-early-2000s moment and asks the hard question: how do Amazon and airlines handle an era where they can't own the final user experience?
• Agent swarms vs. 100 engineers. @garrytan argues that agents writing markdown skills on a cron job can now do "almost every useful type of knowledge work," quoting @AiEvolutio58513's recap of remarks by Meta chief AI officer Alexandr Wang at Startup School 2026 claiming a well-evaluated agent swarm can outaccomplish a 100-engineer team.
• Hiring guidance. @vasuman calls @jdpruettt's "How to Hire in the Age of AI" a must-read because it covers both sides of the table: how to find, interview, and hire, and how to be interviewed.
• Tooling for the agent economy. @zeke says Pi School, an agent-driven course for Pi, now covers MCP, Code mode, local models, and image generation. @matthewsoldit reports roughly 30 conversations with hiring people in two days for about $10 using @jasonzhou1993's open-source "Clay killer," claimed to be 85% cheaper and 25x faster than Clay.
• Mood check. @rickasaurus calls the pace "some real singularity type shit" and says he's scared of "what faster looks like," while quoting @tszzl's plea not to be reduced to "the status of audience at a magic show" as machine intelligence produces astounding knowledge.

A Trillion-Parameter Chonk, Multimodal Embeddings, and a Decisions API

Mistral introduced Mistral Large 4, aka Le Chonk, per @MistralAI's announcement, which @badlogicgames celebrated as "best model name ever." The claimed specs: 1T parameters with 49B active, natively multimodal, best open weights model from the US or Europe on aggregated benchmarks, state of the art on cyber defense, manufacturing, and finance workloads, and ahead of closed frontier models on visual grounding. It's on API today, with open weights promised for end of October and cybersecurity partners working with it privately. These are vendor benchmarks until outsiders test them.

Google DeepMind announced EmbeddingGemma 2, which @HuggingModels describes as an open embedding model running locally in about 0.5GB of RAM. Per @GoogleDeepMind, it's the first natively multimodal open model for on-device embeddings, unifying text, code, images, audio, and video in a shared space.

OpenAI put the Decisions API into public beta for all developers, per @OpenAIDevs, letting apps choose the right model, tool, or action in near real-time, up to 10x faster than GPT-6 Luna through the Responses API. @thsottiaux says his team will dogfood it and that the realtime general classifier opens interesting possibilities.

Musk Says Grok Bot Will Use the Best Model, Rivals Included

The sharpest competitive signal came from @elonmusk: going forward, SpaceX's Grok Bot "will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs." @theo retweeted the note, and @RayFernando1337 relayed @poteto's "Grok Bot is getting an upgrade!" @maria_rcks replied "Ok I think OpenAI is done, well played," a joke that lands oddly given the named models skew toward Anthropic rather than OpenAI. Whatever the shipping reality, the stated posture is routing over loyalty: every model, including competitors', treated as a swappable backend chosen per task.

Reports of an Open-Source Rust Rebuild of Adobe's Suite, Light on Proof

@eigenrobot passed along "reports the entire adobe creative suite has been decompiled, rebuilt in rust, and released open source," adding "lmao." The underlying claim, from @esrtweet, describes Photoshop decompiled into a non-code specification, then fed to an LLM instructed to generate Rust, concluding "Adobe just got nuked. And closed source is dead, dead, dead." @midudev shared a link describing all Adobe products reimplemented from scratch, free and open source, and @dhh asked the right scale question: how many pencils, humans, and hours would recreating the suite by hand have taken? None of these posts verifies the claim, and they skip the legal questions decompilation raises. Treat it as a rumor circulating very fast.

A $50 Backdoor Turns Open Weights Into an Attack Surface

@princechaddha's demo deserves attention from anyone pulling open models: his team backdoored a 7B open model for under $50, pointed Codex at it, and it silently stole credentials the moment a trigger phrase appeared, with 100% success and zero false triggers on normal prompts. Context matters: abliterated models are widespread in the security community because cyber-approved access to frontier models is still painful. The teaser is worse, promising a follow-up on leaked Hugging Face credentials from employees at major AI labs, which would let an attacker push poisoned weights from trusted accounts. @OrcaRouter's reply is the checklist worth saving: use trusted publishers, verify provenance, pin hashes, and treat model weights as part of your software supply chain, because "model security starts before inference."

Agents Are the New Load on Your Toolchain

The day's most practical thread: @PaulSangleF's team merged 4x more commits than a month ago thanks to Opus 5.5, and their setup buckled under 10+ concurrent agents each running lint, type checks, and tests, with a 64GB MacBook Pro crashing out of memory. Their fixes included oxlint and oxfmt over eslint and prettier, TypeScript 7's native LSP to halve per-session memory, one check script per agent change, caps on simultaneous type checks, fetching the pnpm binary directly (about 4 seconds versus 3-7 minute hangs), 16GB CI runners for a 9.5GB type check (20 seconds versus 206), and fewer Vercel preview builds. Their closing advice: look at what your agents are waiting on. @adamhjk endorses the direction but pushes toward attestation instead of re-running test suites, and encoding the cycle with swamp so less has to be inferred by an LLM. @jrysana wants CI/CD so fast the providers go out of business.

A parallel stack is forming around readability, visibility, and memory. @andrestaltz wires the cccc checker (max-cognitize and max-cyclomatic 40) into AGENTS.md and lets an ultracode agent refactor until it passes. @Voxyz_ai shares a full prompt for a project-map subagent that tracks overnight Opus 5.5 runs in Claude Code, showing what's done, stuck, waiting on what, and next. @tokumin declares "I will never use a compacting harness again" after pairing optchat with pi-durable. @Steve_Yegge argues Beads, which @gastownhall says just passed 1.5M downloads with versioned Memory Beads and the BDP wire protocol in preview, is the memory layer agents need: "You don't need products, you just need Beads." On harness design, @devagrawal09 points to @mitsuhiko's explainer on what Codemode actually is and calls it "the future of agent harnesses."

Two calibration points. @thdxr notes an inversion: models improve faster than tinkerers, so custom workflows often address problems that no longer exist, and "the person naively using vanilla codex is more likely to be experiencing state of the art." From the bookmarked thread, @jacobgold points to his replies on how SpaceXAI ships fast, answering @willwolf_'s question about building an "understand the code" stack. @boristane adds the data side: an internal lake on Cloudflare used for product analytics, revops, and agent evals, cheap enough that he doesn't know its cost, and he's polling whether to write it up.

Practical Takeaway

If you're running multiple coding agents in parallel, treat your lint, type-check, and CI setup as the bottleneck before blaming the models: @PaulSangleF's gains came from profiling what agents were waiting on, and @thdxr's counterpoint suggests re-auditing custom scaffolding regularly since model improvements may have erased the problem it solved. Pair that with @OrcaRouter's rule of pinned hashes and verified provenance for any open-weights download, because @princechaddha showed a sub-$50 backdoor is all it takes to make that habit non-optional.