We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Neon on a long optimization task: Theo's numbers
Theo published the most concrete Gemini 4 Neon test so far. The task was his TS Rust project: make the binary smaller and faster. He first gave GPT-6.2 Astra a simple /goal and ran it for 7+ days. Astra made the package 46% smaller but 56% slower, then spent four more days without landing a speedup. Opus and Fable eventually got it about 20% faster after many days . He then forked Neon from the same public starting point with the same structure and goal, and allowed it to use Antigravity's built-in swarm feature . Within 10 hours it delivered:
- a binary 5.95x smaller (47.44 MB down to 7.97 MB)
- up to 20x faster warm starts
- 1.5x faster overall across his benchmark repos (2.3x on Sentry), with full-suite time down 48.6%
The caveat that matters: trust its code more than its explanations. In his architecture analysis of the repo, 10 of Neon's 72 claims were wrong. It still improved that repo the most and the fastest, and Theo concludes it "is verifying its claims when it needs to actually implement" . In his code-review bench it made 21 valid and 5 plausible-but-minor claims, with none wrong . His one reported failure: it didn't invoke custom skills he named explicitly. He couldn't tell whether that came from the model or from Antigravity .
Two of his prompts are worth reusing:
- Loosely bounded setup task. "Get Gemini set up on [machine]," then name the thread to notify. Neon configured Antigravity, added it to T3 Code, verified it, and messaged that thread on his behalf . When the T3 Code adapter hid its HTML output, it rendered the HTML itself, sent screenshots, explained the bug, and filed a PR to fix it when asked .
- Plan as a press release. "Write up an HTML page describing the plan. Write it like a blog post explaining the feature… ignore the style guides from the skill." He says most models ignore that last instruction and Neon did not .
Costs to watch. One user reports that Neon burns Antigravity limits about 3x faster, with a 9% prompt-cache hit rate after 1 minute, versus about 95% for Opus 5.5 and about 97% for GPT in Codex . Theo says he'll go back to pushing Google to fix the platform .
Riley Brown says Neon kept up with Claude Opus 5.5 on many coding tasks. His main complaint is its "It's not this — it's this" phrasing, and he suggests writing a skill to suppress it . He still isn't moving to Antigravity because it has no mobile app (web/PWA remote access is the stopgap) and Google's tool lineup is confusing .
Codex: TUI workspace on by default, larger requests stay on WebSocket
Codex CLI 0.171 turns on the TUI workspace by default. Press Ctrl-X then t for the sidebar. It includes:
-
side chats (
/side) in tabs, saved with the session and restored bycodex resume -
clickable
path:linereferences that open the source at that line -
Markdown preview with
/search and inline images - live subagent status
Charlie Marsh says Codex raised its model-request WebSocket limit from 16 to 128 MB. Large turns no longer fall back to HTTP, and requests that still exceed the limit are about 5x faster .
Side threads and remote ports in other harnesses
-
Antigravity
/btwused to answer one-off questions while the main agent was busy. It now lets you fork, explore and promote side threads without losing your place . That mirrors Codex's/side. -
Antigravity MongoDB Atlas plugin. It connects to MongoDB's hosted MCP server and adds 7 skills, including natural-language queries, slow-query optimization, schema review and Vector Search. Install via
antigravity://customizations/marketplaces/antigravity-plugins-partners/plugins/mongodb-atlas. - T3 Code 0.1.11 tunnels remote environments to localhost. Web apps an agent starts on another machine open in the in-app browser, and any port (such as Postgres 5432) can be forwarded with one click .
Workflow notes
Send annotated screenshots, not PRs. Theo says a picture with arrows and labels lets him, Claude and Neon "understand, agree, and then implement," and he'd take one "over a PR literally any day" .
DHH: no hand-written code, persistent polish. He says agents have built Omarchy for five to six months, producing hundreds of thousands of lines with none written by hand . He argues Linux suits agents because they want CLIs, editable config files and detailed error messages . He hasn't found a setup that one-shots a perfect design; someone still has to spot the problems and keep steering . Cost: most weeks one $200 20x Claude subscription covers everything, he has come close to the limits only about two weeks in three months, and his approach is "Opus 55 fix everything" .
Riley Brown's split stack:
- Claude Code for building. He uses Claude Projects on his iPhone to talk to a coordinator that starts local Claude Code sessions .
- OpenAI for admin and computer use, for example asking Codex to upload a rendered video to YouTube .
- Grok Bot for Twitter research and logging DMs into a sheet .
- Devin Bots on cloud Macs, which built an iOS app, uploaded it to TestFlight and sent back a link; he only had to sign in with his Apple developer account .
Code review is the bottleneck. From a CXO discussion: speed up deterministic checks, and risk review plus auto-merge is workable; Amplitude "got to 81%" (the metric isn't stated) . Kent C. Dodds says his next video covers how they "safely and completely eliminate review of implementation" .
Paper MCP (from a sponsor segment). Paper is a design tool whose canvas is real HTML and CSS. Claude Code, Cursor and Codex can read and write to it, pull in your CSS variables as design tokens, and draw screens with your actual colors and fonts. Once you approve a design, you tell the agent to build it into the app .
- In a TS Rust optimization test, the reviewer started Gemini 4 Neon from the same public baseline and structure as Astra, gave it the same “make it smaller and faster” goal, and enabled Anti-gravity’s built-in swarm; in about 10 hours, Neon reportedly reduced the binary from 47.44 MB to 7.97 MB (5.95× smaller), achieved up to 20× faster warm starts, and delivered a 1.5× aggregate speedup across benchmark repositories (2.3× on Sentry) with 48.6% less full-suite runtime. For comparison, Astra’s first 7+ day run made the binary 46% smaller but 56% slower, then spent four more days without making it faster; Opus and Fable later achieved about a 20% speedup.
- Neon handled a cross-thread setup task: after being asked to prepare a Windows gaming/development PC, it configured Anti-gravity, added and verified it in T3 Code, found the specified thread, and sent a readiness message there. When T3 Code’s Anti-gravity adapter hid HTML-render output for a throughput visualization, Neon diagnosed the missing output, rendered the HTML directly and used screenshots as a fallback; at the user’s request, it then filed a PR for the reported MCP-tool bugs.
- Treat its repository explanations separately from its implementation results: in the reviewer’s TS Rust architecture analysis, 10 of 72 claims were wrong, though the reviewer says Neon delivered improvements when implementing; on a code-review test, it produced 21 valid and five plausible but low-importance claims, with none wrong. It also failed to invoke a specifically requested custom skill, and the reviewer could not determine whether that was a model or Anti-gravity issue.
- DHH says agents have developed his Linux distribution for about five to six months; he has written no code for it by hand and says he has produced hundreds of thousands of lines through agents. He describes prompting changes to the operating system and seeing them take seconds or a few minutes.
- He sees Linux’s command-line interfaces, predictable editable configuration files, and detailed error messages as strengths for agents, which can use these surfaces to inspect and change a system.
- His workflow still depends on human judgment: he spots imperfections and steers agents through repeated refinements. He says he has not found a setup that produces perfect designs in one shot, and emphasizes persistence through the final polish rather than stopping at a working first pass.
- On usage, he said one $200 “20x” subscription was enough most weeks, he rarely came close to its limit, and he was using Opus to “fix everything.”
- LangSmith’s coding-agent view connects GitHub PR outcomes with coding-agent spend and traces, reports metrics such as cost per merged PR by model, user, or PR type, and classifies threads such as bug fixes versus exploration. The listed harness integrations include Pi, OpenCode, and Codex.
- A practical improvement loop is to use traces as the record of agent behavior, run offline evaluations before changes and online evaluations for A/B tests or regression checks, then turn failures observed in production into evaluation cases and regression/unit tests.
- LangSmith Engine monitors traces for error patterns and creates evaluations and code fixes; Chase describes it as a coding agent equipped with specific skills, tools, and CLIs, with customization as a focus.
- LangChain Deep Agents is an open-source, model-agnostic base harness intended for modification through middleware or a fork. It includes filesystem, sandbox, and browser tools; skill loading, summarization and context offloading, prompt caching, human-in-the-loop support, and subagent/async delegation. Chase describes it as a base harness, not specifically an end-user coding harness, though it is capable at coding.
- For coding and making things, he primarily uses Claude Code; he can start local sessions from his iPhone by talking to a coordinator, and uses a main Claude Project for broader work such as decks and documents.
- For iOS builds, he used a Devonbot cloud-Mac session to create an app, upload it to TestFlight, and send back a download link; he needed to sign in with his Apple developer account, but did not need to run the app on his own Mac.
- He routes admin work to OpenAI, including email and Slack replies and computer/browser tasks (for example, asking Codex to upload a video after rendering); he uses Grokbot for Twitter research and sorting DM information into a sheet.
- In his tests, Gemini 4 Neon handled coding tasks close to Claude Opus 5.5, but he had not switched to Antigravity: it lacked a mobile app, with web/PWA remote access offered as an interim option, and he found Google's tool choices unclear.
Sanfilippo argues that developers should use AI coding to rethink what software can do, not just implement existing plans faster: agents can make narrow, purpose-built applications viable to build, potentially in three days, and speed up product and interface iteration. He is testing this approach with Grano, a Redis fork that extends compression and encoding ideas beyond what he previously considered a sensible effort. As an example, he cites CNC Kitchen’s Stefan having Claude remake software for a Canon scanner; Sanfilippo says the result was better than Canon’s software and took one afternoon.
- LangSmith’s coding-agent view links traces and spend to GitHub PR outcomes, with metrics such as cost per merged PR and cost by model; it can group by user, model, or PR type and classify threads as bug fixes or exploration. The talk lists integrations for Claude Code, Pi, OpenCode, and Codex.
- A reusable improvement loop is to identify production failures in traces, turn them into eval cases, and run them as regression tests; use offline evals before changes and online evals for A/B testing or regression checks. LangSmith Engine monitors traces and generates evals and code fixes; the talk describes it as a coding agent with skills, tools, and CLIs, with people reviewing and merging its changes.
- LangSmith’s Tool Gateway can govern coding agents across frameworks: it tracks which agent calls which tool on whose behalf, applies identity and policy checks, and logs calls for audit; allowed requests can be forwarded to hosted tools, remote MCP servers, or self-hosted tool servers.
Gemini 4’s comparison graphic positions Carbon for the hardest long-horizon projects and deep research, Neon for deep reasoning across software engineering, knowledge work, and science, and Hydrogen for speed and efficiency in everyday coding and agentic workflows; its listed quality-index scores are 60, 61, and 54 respectively, with speed ratings of 2, 3.5, and 5 out of 5 . The graphic lists API prices per 1M input/output tokens as Carbon $4/$20 (promotional $2/$10 through Nov. 30), Neon $3/$15, and Hydrogen $1.50/$7.50 . Simon Willison asks whether Argon is effectively retired because it is omitted from the comparison graphic; the post does not establish that it has been retired .
- Brown reports using Claude Code for coding and creation, with Claude Projects on iPhone as a way to talk to a coordinator and spin out local Claude Code sessions; he routes admin/email and computer-use tasks to OpenAI, and Twitter research, DMs, and logging opportunities to Grokbot.
- Devon Bots’ cloud-Mac workflow let Brown ask an agent to build an iOS app, upload it to TestFlight, and return a download link; setup required signing in with an Apple Developer account.
- Brown recommends using agent platforms that can act on a computer rather than relying on chat alone. He says Gemini 4 Neon handled his coding tasks well, close to Opus 5.5 on many tasks, but he was not switching to Antigravity because it lacked a native mobile app (web/PWA remote access was the interim option) and Google’s tool lineup left him unsure where to commit.
Keshav says he built four projects with AI over 14 months: a local Wispr Flow alternative with paying users, an AI newsletter-writing interface, a self-serve sponsorship platform, and an app based on a board game.
- The creator used Soul 6.2 in ultra-fast mode with one prompt to build a streaming-chat app for Twitch, Kick, Google, and X; set up credentials and environment variables; deploy it on Cloudflare; test sign-in and chat receipt locally; add WorkOS email access with an approved-account allowlist; and check it inside OBS’s embedded Chrome. They reported that it worked after 45 minutes, but did not inspect the code and intended to rewrite it—so this demonstrates rapid integration and prototyping, not verified code quality.
- To judge a model, try it on real tasks in your own domain instead of relying on online reactions or scores: the creator notes that evaluations may not match your problem area and may not disclose what was tested, and recommends giving a model a sustained trial and noting where it succeeds or fails.
Kent C. Dodds previews a video on safely and completely eliminating implementation review. The linked code-review discussion recommends speeding up deterministic checks and describes risk review plus auto-merge, saying Amplitude “got to 81%” without specifying the metric; it also says teams that get this right improve velocity and quality.
Cursor/Graphite’s Origin was announced as a Git competitor built to scale for agent workloads, extensible via API and MCP, with built-in merge-conflict and agent-assisted co-failure resolution . In a subsequent post, swyx asked whether developers were using it in practice and whether it had Actions or CI/CD; those questions leave adoption and those capabilities unconfirmed here .
Paper’s MCP server lets Claude Code, Cursor, and Codex read and write to an HTML/CSS canvas; an agent can import CSS variables as design tokens and draft a screen using the app’s colors and fonts, after which a developer can adjust it and ask the agent to implement the approved design. The video says this avoids a translation layer between the design and shipped UI. Paper is identified as being used in production by teams at RAMP, Lovable, and Linear.
ThePrimeagen reports that further testing of Google’s Neon showed promise on engineering tasks, including metaprogramming, which he says it handled wonderfully .
Riley says he tested Gemini 4 Neon for 72 hours and found it very good, while raising concerns about its writing ability; he also discusses using it in Antigravity and what is keeping him from committing to that environment. The video also covers Devin Bots, GrokBot updates, and his current AI stack, though the post gives no details about those tools or workflows.
Kent C. Dodds recommends trying Kody (kody.codes). In the post he quoted, @jjenzz says setup was easy and describes Kody as their inbox, agent trigger, sanitizer, credentials manager, and hosted dashboard.
MaxBlade described an open-source browser battle royale as “100% vibe coded,” with hundreds of pull requests and billions of tokens invested in making the game. The game runs in a browser without a game engine; the author reported 40 FPS on a 2021 M1 Mac.
Riley Brown predicts Claude and Codex will increasingly be able to rent compute on demand, potentially raising GPU prices; as an example, he says a heavy Blender or other 3D render could take days on a MacBook versus 30 minutes with agent-accessible compute. This is a proposed agent workflow, not a report of an existing capability.
LangChain included LangSmith for Coding Agents in a weekly announcement roundup, alongside the LangSmith Tool Gateway, a new Agent Platform, and controls for sensitive data in traces; the post gives no product capabilities or usage details.
- LangSmith’s coding-agent analytics is agent-agnostic: it ties agent spend to produced pull requests, shows the cost of merged PRs versus work that did not ship, and helps investigate costly PRs by model, skill, or runaway thread. Teams can then change a model or policy and track whether cost per merged PR improves.
- LangSmith Tool Gateway offers a reusable MCP access pattern: register MCP servers once, group tools into a kit, and provide agents or MCP clients one URL. It checks the caller on each request, attaches the appropriate user or agent-profile credentials, and logs calls; agents do not handle tokens, and monitoring, tracing, and auditing are available in the same workspace.
- For trace access, teams can let engineers use metadata while restricting inputs and outputs; permissions can be set by user group and trace metadata to maintain least-privilege access.
- In a TS Rust optimization test, the reviewer started Gemini 4 Neon from the same public baseline and structure as Astra, gave it the same “make it smaller and faster” goal, and enabled Anti-gravity’s built-in swarm; in about 10 hours, Neon reportedly reduced the binary from 47.44 MB to 7.97 MB (5.95× smaller), achieved up to 20× faster warm starts, and delivered a 1.5× aggregate speedup across benchmark repositories (2.3× on Sentry) with 48.6% less full-suite runtime. For comparison, Astra’s first 7+ day run made the binary 46% smaller but 56% slower, then spent four more days without making it faster; Opus and Fable later achieved about a 20% speedup.
- Neon handled a cross-thread setup task: after being asked to prepare a Windows gaming/development PC, it configured Anti-gravity, added and verified it in T3 Code, found the specified thread, and sent a readiness message there. When T3 Code’s Anti-gravity adapter hid HTML-render output for a throughput visualization, Neon diagnosed the missing output, rendered the HTML directly and used screenshots as a fallback; at the user’s request, it then filed a PR for the reported MCP-tool bugs.
- Treat its repository explanations separately from its implementation results: in the reviewer’s TS Rust architecture analysis, 10 of 72 claims were wrong, though the reviewer says Neon delivered improvements when implementing; on a code-review test, it produced 21 valid and five plausible but low-importance claims, with none wrong. It also failed to invoke a specifically requested custom skill, and the reviewer could not determine whether that was a model or Anti-gravity issue.