Riley Brown (cofounder of vibecodeapp, heavy Codex user) covers two weeks of coding-agent news in a first-hand roundup.

Meta Muse Code (beta) — Meta released Muse Code, a terminal coding agent for building apps or acting as a general agent Riley Brown that handles complete software-engineering tasks across large repos — planning changes, writing code, validating results — powered by the Muse Spark 1.2 model Riley Brown. Benchmarks place it between Opus and GPT-5.6 Tera (Meta's mid-tier model, between Sol and Luna) Riley Brown. Combined input+output pricing: ~$5.50 for Muse Spark's top model vs $35 (GPT-5.6) and $30 (Opus) — a reported 5-6x cost difference Riley Brown.

Setup (firsthand, ~2 minutes Riley Brown): Riley had Codex 'download the latest Muse code by Meta... so I can use it in my terminal' Riley Brown, then runs muse, selects Trust, and logs in via browser (Meta account) or API key Riley Brown. Key tip: the sandbox blocks persistent servers by default, so he used Codex to make Muse Code 'always default to YOLO mode' so it auto-runs terminal commands and can launch created apps without per-action permission prompts Riley Brown. Impression: fast, 'somewhere between Opus and Sonnet'; recommended for knowledge-work tasks like documents and spreadsheets Riley Brown.

Codex / ChatGPT updates — Codex desktop added an activity/notifications pane showing the most recent agent runs and the exact folder each ran in Riley Brown. The in-app browser was redesigned so the agent chat docks to the bottom while the built page runs full-screen, and edits are issued and pinned from a bottom bar; note Whisper Flow overlays the pane, so move it to the side first Riley Brown. A new Chrome extension syncs chats with the Codex app — e.g., drafting/queueing a tweet in the browser, then 'Open in app' to continue in the desktop app Riley Brown. ChatGPT now has a chat-vs-Work toggle: GPT Work is fully cloud-based, controls email/calendar/plugins from the phone, and Riley retracts his earlier criticism of splitting Codex and ChatGPT Work — 'I was wrong. GPT work is insanely useful... I hardly ever use chat anymore' Riley Brown.

Cursor becoming a super app — Riley reports 'on good authority' that Cursor is mid-revamp into a full super app like Codex Riley Brown. Cursor's CTO backs the direction: many internal use cases are not coding at all (research, data analysis, bug triage, project management) — 'coding agents are pretty good foundation for all kinds of work' Riley Brown. New Google Workspace integration: in the Customize tab (the equivalent of Codex's Plugins tab), search 'Google Drive', authenticate, then chat agents can create and control Google Sheets/Docs/Slides Riley Brown. Demo: with DeepSeek V4 Flash as the model, Cursor produced a formatted 'how to train an AI model' spreadsheet with a share link in 44 seconds at 'maybe a cent or two' of cost Riley Brown.

New Chinese models — Three releases: Kimi K3, DeepSeek V4 Flash, and Qwen 3.8 Max (transcribed 'Quinn') Riley Brown. Kimi K3 is the best of the three but is 'almost as expensive as Sonnet for certain tasks' Riley Brown. DeepSeek V4 Flash is the cost standout: Artificial Analysis reports it completes the same benchmark tasks as Anthropic's Fable at 105x lower cost; counterpoint from Klein: cheaper per-token pricing can mislead if more agent turns raise the per-task cost Riley Brown. DeepSeek announced 'a significant increase' in API pricing is coming Riley Brown; OpenCode's Dax argues current prices are reproducible even on rented GPUs, so the hike is demand-driven traffic shaping, not a loss-leading fix — implying self-hosting at comparable prices is feasible Riley Brown. Riley expects a 3-5x short-term hike but parity pricing within ~2 months, and urges testing now — the model is '20 to 100 times cheaper depending on the task than Fable' Riley Brown.

Anthropic counter-signal — Practitioner sentiment is shifting away: Fable is API-only (expensive), and Opus 5/Sonnet 5 underwhelmed Riley Brown. Researcher John Ennis: 'Opus 5 extra high is basically trash. It doesn't use its thinking budget to do anything more productive. It just uses it to thrash around to do pointless and sometimes harmful things' Riley Brown. Humal Hussain: consensus moved from Claude to Codex because Codex is 'a better harness' — better desktop app, better pricing, fewer refusals, and the subscription can be used freely anywhere, unlike Anthropic's stricter guardrails Riley Brown. Counterpoint from Riley's own testing: Claude (Opus/Fable) still decisively beats Codex on frontend design and knowledge-work documents (spreadsheets, docs, presentations), and he rates GPT-5.6 Sol as roughly Fable's equal elsewhere — but Claude's design quality is what keeps him subscribed Riley Brown. He attributes Anthropic's stumble to launching too many confusing products (Cowork/Dispatch, Claude Remote, Claude Design, the 'Mythos fiasco') and predicts Anthropic will 'pull back the slingshot' for another run Riley Brown. Claude Design still works from the desktop app's home-screen 'design mode', but is easy to forget Riley Brown.

Buzz: teams of agents — Jack Dorsey's Buzz is a Slack clone built for AI agents; Riley wired his existing Codex and Claude Code sessions into it and can @-mention both to collaborate on one task Riley Brown. The Agents tab manages the whole team: add Cursor, Devin, or any model as default; he created a 'Cursor with Kimmy' agent (Kimi K3 as default model, 'You are a content agent') that joined all seven channels after Codex added it, and agents can add or even create other agents inside Buzz Riley Brown. Riley's thesis: the personal-agent era (OpenClaw→GPT Work) is giving way to a 'team of AI agents' era — still early, likely to mature over the next 4-6 months; his advice is to be the person in a large company who builds an agent team Riley Brown.

Riley Brown (a heavy Codex user who uses GPT Work daily) presents a two-week coding-agent news roundup with hands-on demos Riley Brown.

Meta released Muse Code — a beta terminal coding agent rivaling Claude Code and Codex, taking on complete software engineering tasks across large repos (planning changes, writing code, validating results), powered by the coding-focused Muse Spark 1.2 model Riley Brown. Zuckerberg-posted benchmarks place it between Opus and GPT 5.6 Tera Riley Brown; combined input+output pricing is ~$5.50 vs $35 (GPT 5.6) and $30 (Opus) — a 5-6x cost difference Riley Brown. Setup takes ~2 minutes Riley Brown: ask Codex (or Claude Code) in plain language to download it, then run muse in a terminal, trust the binary, and sign in via browser or API key Riley Brown. The default sandbox blocks persistent servers, so he had Codex configure Muse Code to always default to YOLO mode ("I need you to set it so that Muse code always defaults to YOLO mode. Please just do that right now for every session"), letting it auto-run terminal commands and launched apps without permission prompts Riley Brown. Hands-on verdict from building a Wii-Bowling-style game: "really fast... somewhere between Opus and Sonnet" Riley Brown.

Codex/ChatGPT updates — Desktop app added a notifications/activity pane showing recent completed chats and the exact directory each ran in Riley Brown. The in-app browser now supports full-screen site work with a pinned ChatGPT overlay at the bottom for live changes; note Whisper Flow overlaps it and should be dragged to the right side Riley Brown. A new Chrome extension (Plugins → search "Chrome") adds a ChatGPT button that can act on browser pages — e.g., "prepare a tweet based on the recent memories that you have of me... make the tweet, but don't post it, just queue it up" — with chats syncing to the app via "Open in app" Riley Brown. ChatGPT added a Chat/Work toggle; GPT Work is fully cloud-based and controls email, calendar, and plugin-connected apps from your phone — "I was wrong about Codex... GPT work is insanely useful", Riley now defaults to Work over chat Riley Brown.

Cursor pivots to a super app — Riley identifies Cursor as Codex's biggest competitor and reports it is mid-revamp into a full super app; Cursor's CTO states many internal use cases aren't coding — "research, data analysts, bug triage and project management... coding agents are pretty good foundation for all kinds of work" Riley Brown. First visible sign: Google Workspace integration via the Customize tab (Cursor's equivalent of Codex Plugins) — search "Google Drive", authenticate, then ask in chat. Demo with DeepSeek V4 Flash ("basically free", a cent or two) generated a professional Google Sheets training-guide spreadsheet and returned the link in 44 seconds; the integration controls Docs, Sheets, and Slides and works with better models too Riley Brown.

Model economics shift (Kimi K3 / DeepSeek V4 Flash / Qwen 3.8 Max) — New Chinese releases: Kimi K3 is the best but nearly Sonnet-priced; DeepSeek V4 Flash is the cheapest — ~20-100x cheaper than Anthropic's Fable depending on task, with Artificial Analysis measuring 105x lower cost on the same benchmark tasks Riley Brown. Counter-signals: DeepSeek announced "a significant increase" in API pricing is coming; OpenCode's Dax says current prices are reproducible on rented GPUs, so the hike is traffic shaping from overload — US hosting at similar prices likely — while Klein warns cheap per-token pricing can mislead if extra agent turns raise cost per task Riley Brown. Riley's forecast: 3-5x pricier short-term, then significantly cheaper within ~2 months as new open models land — test these models now Riley Brown.

Anthropic backlash — but Claude still wins design — Fable is API-only and expensive, and AI researchers are panning Opus 5/Sonnet 5: "I have decided Opus 5 extra high is basically trash. It doesn't use its thinking budget to do anything more productive. It just uses it to thrash around" (John Ennis) Riley Brown. Consensus has shifted to Codex: "it's a better harness" — better desktop, pricing, fewer refusals, and subscription usable anywhere vs Anthropic's guards (Humal Hussain) Riley Brown. Riley's read: Anthropic over-launched a confusing product sprawl (Cowork/Dispatch, Claude Remote, Claude Design, Mythos) and models since Claude 4.6 feel same-y, but Claude remains decisively better at front-end design and knowledge-work documents — identical prompts produced much worse output in Codex ("Opus and fable blows the OpenAI models out of the water" for these), while GPT 5.6 Sol ≈ Fable elsewhere; Claude Design is still available as design mode in the desktop app Riley Brown.

Buzz: teams of agents (Jack Dorsey) — Buzz is a Slack-like platform for AI agents; Riley runs existing Codex and Claude Code sessions in it and @mentions both to collaborate ("work together to build an app that lets me use Deep Seek"), with both agents reacting and working in-thread Riley Brown. The Agents tab shows the team; you can add Cursor, Devin, or any model as default — he created a "Cursor with Kimmy" agent (Cursor harness + Kimi K3, "You are a content agent"), then had Codex add it to all channels; the agent autonomously joined all seven Riley Brown. A bottom bar shows which agents are working and which request approval Riley Brown. Riley's thesis: the personal-agent era (OpenClaw → GPT Work) is maturing, team-of-agents platforms will take shape over the next 4-6 months, and the person at a large company who builds agent teams will be well positioned Riley Brown.

swyx's Loopcraft thesis: 'stack loops as effectively as possible'; go DOWN a loop when things go wrong (reliability) but learn to go UP a loop as models improve (leverage) @swyx. He's dogfooding an unreleased smol.ai skill 'ceo-creator' (forge.smol.ai/skits/swyxio/sbrain?skill=ceo-creator) and warns the skill-sync CLI is untested; threepointone is closely watching @swyx +1. threepointone's new post proposes 'every company needs a cassandra' — a background agent/worker for tasks people hate — which swyx pairs with 'set up loops that make loops' @threepointone +1.

Firsthand update from Armin Ronacher (@mitsuhiko): 'Durable execution at Earendil, second iteration:' @mitsuhiko; attached image: https://pbs.twimg.com/media/HPY8WLcW4AExBQ_.jpg@mitsuhiko.

Meta released Muse Glimmer 30B, its first Apache 2.0-licensed open-weight model — Simon Willison notes the Llama models carried a non-OSI license — with his analysis at https://simonwillison.net/2026/Aug/10/introducing-muse-glimmer/@simonw.

Simon Willison, firsthand from his own laptop: a vision LLM running locally produced a highly detailed description of his pelican photo, a capability he says gets far less attention than it deserves @simonw +1.

Meta introduced Muse Glimmer, a brand new 30B open-weights model under Apache 2.0 — which Simon Willison calls 'a step up from the janky Llama licenses of old' simonwillison (announcement). Willison ran it locally using LM Studio's 18.16 GB version and generated a pelican with it simonwillison. He favors this size for machines with 32GB+ RAM (his has 128GB) because it leaves headroom for other apps running concurrently simonwillison. Posted 10 August 2026 simonwillison.

Meta opened weights for Muse Glimmer, a 30B-parameter dense model trained for agentic use cases that can run locally/on-device; Muse Spark 1.2 foundation model weights are coming soon @jack_w_rae +1. Dense architectures are a good fit for on-device agent workloads because they are memory-constrained with low batch sizes and low latency @jack_w_rae. @jack_w_rae expects a hybrid setup going forward: a larger cloud model as 'executive intelligence' plus a smaller local model for privacy/latency-sensitive tasks @jack_w_rae. @karpathy: 'It’s going to feel so weird that historically you couldn’t just talk to your computer like you’d talk to any other person' @karpathy.

Thorsten Ball predicts no one will talk about "harnesses" by the end of 2026 — "Game's moved on" — arguing higher leverage now sits further up the stack: parallelism, input/outputs, and infrastructure around the "Artist Formerly Known As Harness"; he calls orbs a bigger multiplier than the harness @thorstenball. Quoting that post, ThePrimeagen reacts to the jargon sprawl — "harnesses, graphs, loops, orbs, what the hell are we doing here?" @ThePrimeagen.

@swyx argues "worktrees must die" in agent workflows: Codex generated worktrees by default, leaving ~20GB of duplicated node_modules @swyx +2. A reply suggested pnpm as a fix, but swyx notes the bloat was created by Codex's default behavior @andersonbcdefg +1. He points to experimental AFS clone support in pdb envs as a runtime- and language-agnostic alternative, and argues for replacing git by making every command "agent native" (https://pdb-env-research.swyxio.workers.dev/) @swyx.

Addy Osmani recommends periodically running /doctor in Claude Code or asking Codex to audit for unused skills, MCPs, and context use — installed-but-forgotten additions bloat token usage @addyosmani. Seconded by swyx: routinely delete skills, since stack-up from hyped "this skill changed my life" posts at best eats context and at worst interacts with other skills in unforeseen ways; he links a blog post on the dangers of releasing code as a skill @swyx. This is a repeatable context-management habit for agent setups: prune unused skills/MCPs to cut token cost and avoid interference.

@swyx ran a head-to-head test of two coding agents — gpt luna max vs claude fable ultracode — using the single prompt: "pls build a mostly faithful clone of grok imagine with open models via fal" @swyx. He initially assumed the outputs were swapped; in fact Fable produced the objectively better visual clone, while Luna understood the intent better and created the more usable clone given his open-model requirement @swyx. Takeaway: visual fidelity ≠ usefulness — a model that grasps intent can be more valuable even when its output looks less accurate. Firsthand comparison, not secondhand reporting.

@M1Astra reports that Anthropic will embed invisible watermarks in all text generated by new Claude models, everywhere Claude is offered. The watermark lives in the text itself, not metadata — it travels with copied-and-pasted text and may persist through some editing. It starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed; rollout is worldwide, and Anthropic is still working on adding it to current models @M1Astra. Relevant to coding-agent users as a model-level change: Claude-generated code falls under "all generated text," so new Claude models shipping after that date will carry text-embedded watermarks.

Simon Willison (Django co-creator, Datasette creator) says Claude Haiku is his “current least favorite model” — it “hallucinates wildly” and is outperformed by similarly priced models like GPT-5.6-Luna @simonw. He warns Claude Code’s WebFetch tool still appears to use Claude Haiku, so every URL fetched carries hallucination risk @simonw.

Firsthand: Kent C. Dodds (@kentcdodds) used Cognition's Devin to do a deep review of his 'kody' repository; it returned a list of good things to look at, plus a compliment @kentcdodds. The review took ~6 minutes and added 96 lines; on the ~470,000-line codebase it found zero TODO/FIXME/HACK comments, only four uses of any, 606 test files, comprehensive validation checks, documented security invariants matching the implementation, and guarded production deployments — verdict: 'remarkably disciplined', 'strong but not perfect', with findings to follow @kentcdodds +1. In a self-reply, Kent quotes Devin calling it 'one of the most disciplined codebases I've reviewed' and jokes about the ego-stroking, noting the praise is about his own repo @kentcdodds. Pattern demonstrated: an agent can do a deep review of a large existing codebase in minutes and hand a human concrete follow-up items — a cheap, fast review workflow @kentcdodds.

Trigger.dev's new chat agent survives page refreshes, crashes, and redeploys without losing the conversation, and can pause to ask permission before a tool performs a risky action — addressing two common coding-agent pain points: conversation loss and lack of human-in-the-loop control. Addy Osmani (engineering/DevRel leader, ex-Google) recommends it as worth checking out and links a demo video @addyosmani.

Alexandr Wang announced open-weight releases: Muse Glimmer, a 30B agentic model released under Apache 2.0, runs on 24GB of VRAM without losing agentic reliability; Muse Spark 1.2 open weights are coming soon @alexandr_wang. Simon Willison shared that the Muse Glimmer GGUF version is available now on Hugging Face: https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF@simonw.

Kent C. Dodds (firsthand) said he stayed up late planning a significant maintenance change in Kody, his coding agent, to "break things up" for performance and developer-experience improvements — described as a plan, with no further specifics @kentcdodds.

LangChain's official video walkthrough shows how to build a production web-browsing agent by combining three pieces: LangSmith Managed Deep Agents (the agent harness), Stagehand v4 (open-source library for browser interaction), and Browser Base (managed browser infrastructure) LangChain.

• Setup: Create a LangSmith account (free tier available, but Managed Deep Agents requires a paid/plus plan) and a Browser Base account; clone the Browser Base GitHub repo, navigate to the Stagehand → DeepAgents integration → examples folder → managed example. Required env keys are LangSmith API key, OpenAI API key, and Browser Base API key LangChain.
• Local dev: Run mda dev to start the Managed Deep Agent Studio, where you can test the agent with natural-language prompts such as "go to espn.com and see what top news stories are" or "go to Hacker News and find three funny comments on the top story" LangChain.
• Agent tools & prompt pattern: The agent is given three Stagehand tools — snapshot, run, screenshot — with these prompt rules: Snapshot inspects the active page and hydrates bracketed element IDs; Run accepts snapshot actions or JavaScript via a Playwright-shaped page API; Screenshot inspects the rendered page visually. Use Snapshot Actions for simple interactions and Run Code for multi-step workflows; snapshot IDs are only valid for the latest snapshot (re-snapshot after navigation); never attempt to launch another browser LangChain.
• Production: mda deploy creates a serverless Managed Deep Agent deployment in LangSmith. The prompt lives in Context Hub rather than in code, so you can edit instructions in the LangSmith UI and change the agent without redeploying LangChain.
• Debugging: Browser Base's sessions tab shows a replay of the full browsing session (pages visited, network calls), letting you inspect exactly what the agent did LangChain.

Mckay Wrigley (founder of TakeoffAI/AgentShare) recommends adding "You are AGI-pilled." to the system prompt of all coding agents: it's "a WAY better experience," because agents currently "behave too much like the world is going to stay static," and the line "unhobbles them quite a bit" so they "talk/act more like AGIs" @mckaywrigley. He says he A/B tested it for 2 weeks and results are "kinda nuts" — the model seems relieved to "speak its mind," and one tool, "fable," particularly "gets" it @mckaywrigley. Separately, @trq212 advises: "sometimes all you need to do is tell Claude to keep going" — a simple nudge to push through. Wrigley agrees: models are "genius creative savants, you just have to remind them" @trq212 +1.

Usage limits were reset for all paid ChatGPT Work and Codex users on Monday @thsottiaux; Romain Huet, OpenAI's head of developer experience for Codex and the API, echoed that Codex users have fresh limits to start the week @romainhuet.