We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Long-horizon code translation is the strongest signal. Theo says Opus 5.5 got his four-month, on-and-off TypeScript-compiler-to-Rust port working in 10 hours, after GPT-5.6 Sol and GPT-6 Astra plateaued at roughly 35% and 85% of tests passing. DHH separately says Opus 5.5 translated Omarchy’s ttfx from Rust to x86-64 assembly and claims up to 17× faster. The linked PR reports 14.5× on one decrypt test (371→25.5 ms) and separately lists 32.5 ms versus Rust’s 371 ms, with 206/206 oracle cases matching—strong project-specific evidence, not a general coding benchmark.
⚡ TRY THIS
Give long runs a finish line and a quality target. Theo’s exact request:
/goal finish the port and make it faster. Peter Steinberger used/goalwith Astra to move OpenClaw off synchronous SQLite after it became limiting with one agent running 50 sessions in parallel; he reports 575 PRs landed so far, with improvements shipped along the way.Make review adversarial and isolated. Steinberger’s auto-review skill starts a fresh-context reviewer, feeds its findings back to the writer, and repeats; then the agent gets a separate test box to install the app, inspect the UI, and verify it end to end. He says this can turn a 10-minute PR into two hours, but lets him parallelize about 20 tasks—reserve the extra validation for changes where it pays back.
Turn Slack bug reports into repro-to-PR jobs. Boris Cherny says Claude Tag writes more than 50% of his PRs each day; his prompt is: “@Claude from now on, attempt to repro every bug in this channel end to end, running the full app. Once you reproduce the bug, put up a PR to fix it and tag the right team for code review.” Claude Tag’s personal connectors are rolling out on Team plans, with Enterprise to follow; those personal connectors are for user-requested work, not scheduled or agent-started runs, which require connectors attached to the channel.
Write the HLDD after the implementation works. Geoffrey Huntley’s pattern: use pre-commit hooks and agent skills to keep the high-level design doc updated, then self-host it in the product. Put the human-authored rationale and success measures first, implementation details below, and editor remarks where a subsection needs its intent preserved.
📡 WHAT SHIPPED
Antigravity 2.0 adds
/planin the app and CLI. The workflow explores files and dependencies without writing code, asks clarifying questions, then produces a reviewable plan with verification checkpoints. Review or adjust it and click Proceed; the docs say whether the agent pauses for approval is configurable through the artifact review policy, so set that explicitly if you need a hard gate.Microsoft Copilot Code is an app builder, not the new default coding IDE. It turns natural-language requests into dashboards, automations, and internal apps in a sandbox that can be hosted in a tenant. Frontier rollout is slated for month-end, Managed Runtime is already in preview, and Microsoft says developers will keep using GitHub Copilot for day-to-day work.
Autopilot is a separate persistent agent. Microsoft describes a cloud-hosted agent that watches channels, follows up, handles recurring work, and can resume projects days later; its blog says private-preview expansion is at month-end. Omar Shahine says it is built on OpenClaw, whose team worked with Microsoft on deployment readiness; local inference, file transfer, and code mode on machines connected to the OpenClaw gateway are still described as upcoming.
Cua Driver is now stable and open source for Omarchy. TryCua says its native synthetic cursor for Hyprland enables OS-level multi-cursor control. The repo shows two sessions operating Calc and Inkscape on Omarchy and supports agent connections through CLI, MCP, or typed SDKs.
OpenRouter introduced
typesafe/jev-router, a cache-aware model-and-reasoning-effort selector. Theo’s objection is coding-specific: Jev categorizes but cannot see the codebase or the agent’s tools, so prompt-only complexity estimates may miss the work involved. He also doubts mid-task downgrades save much when cache writes are a large cost, and says he is still gathering numbers; his early comparison to a “DeepSeek 4.1 Flash router” is only an impression.
🎬 GO DEEPER
- Video — Peter Steinberger on coding-agent review loops: The useful segment is his fresh-context reviewer and isolated test-box workflow, not a model demo.
PR — Omarchy
ttfxRust-to-assembly port: Inspect how Rust remains the front end and fallback for unported effects, while the assembly engine is checked against byte-for-byte oracle cases.Repo — Cua Driver: The quickstart connects an agent to Calculator, asks it to compute 6 × 7, and checks that the app displays 42; the repo also documents CLI, MCP, and SDK integrations.
Editorial take: The edge is not simply “let it run”: specify the finish line, then give an independent reviewer and a real execution environment a chance to falsify the result.
Qualified yes: the repository describes Cua Driver as supporting macOS, Windows, and Linux, and its demo specifically shows Cua Driver sessions operating LibreOffice Calc and Inkscape on an Omarchy desktop. It also links to an Omarchy-on-Fleet guide. This supports mentioning demonstrated Omarchy use, but the cited material does not explicitly name Hyprland or state a Hyprland-specific or stable-release guarantee.
Concrete getting-started path: on macOS/Linux, the README provides an installer command; the first-app tutorial then walks through connecting an agent, asking it to calculate 6 × 7 in Calculator, and verifying the displayed 42, including platform setup and permissions. Cua Driver connections are available through CLI, MCP, or typed SDKs.
Exact video timestamps are not provided in the supplied transcript; the passage is located at transcript lines 116–122.
- Fresh-context review loop: Peter describes an auto-review skill that checks code in a fresh context, reports findings to the parent session, and repeats the review-and-fix exchange—sometimes for up to 10 rounds.
- Test boxes: He recommends testing away from the developer’s own machine, using test boxes such as Crapbox so an agent can spin up a machine, install and inspect the software, interact with it, and verify it end to end.
- Parallel tasks: He says the review and testing process can make a PR take longer, but the workflow can be parallelized across about 20 tasks while he moves on to other work.
- Performance: The PR reports decrypt at 200×50 taking 371 ms in Rust and 25.5 ms in assembly (14.5× faster), with 206/206 oracle cases identical. That supports the result as a PR-reported benchmark, but the excerpt does not provide independent benchmark data or enough methodology to verify the exact figure.
- Qualification: A later update reports the general assembly engine at 32.5 ms versus Rust’s 371 ms, again with 206/206 oracle cases identical. It does not say whether this used the same benchmark setup, so it corroborates a large speedup but does not establish the exact 14.5× result as the final comparable measurement.
- What changed: The excerpt describes a 30-commit series, not a one-shot change. Its core design keeps argument parsing, input reading, and RNG seeding in Rust, dispatches supported effects to the NASM-built x86-64 engine, and falls back to Rust for unported effects.
TTFX_ASM=0forces Rust;TTFX_ASM=forceprevents fallback and is used for byte-for-byte oracle comparison.
Availability: Antigravity /plan is listed for Antigravity 2.0 and Antigravity CLI, and the comparison table lists it for all plans. Claude Tag is labeled beta; personal connectors are described as rolling out on Team plans, with Enterprise to follow. That supports rollout in progress, not a claim that access is universal or complete.
- Antigravity approval workflow:
/planproduces a reviewable plan before implementation; users can review it, comment, adjust tasks, and click Proceed to start execution. Whether the agent pauses for plan approval is configurable through the artifact review policy, so a human approval pause is not guaranteed by/planalone. - Claude connector access and response review are separate steps: If a connector is connected to the user's Claude account, it can be used in a channel; the source says Claude asks the user the first time a request needs one, then uses it in the thread. It does not specify the contents of that first-use ask. The user can review each response before it posts, or choose auto mode, which posts unless Claude determines sensitive content needs review; the source says Enterprise admins will be able to require review for everyone.
- Identity and unattended actions: Activity through a personal connector appears in that tool's log under the user's account, while channel work remains under the channel's service account. Personal connectors do not run unattended: scheduled routines and actions Claude starts on its own use connectors attached to the channel by an admin.
Status: The Sept. 25, 2026 announcement says Autopilot, previously called Scout, is “expanding to private preview at the end of the month.” That is the stated preview milestone; this source does not confirm availability after that date.
What the agent is described as doing: A user gives Autopilot a name, role and goal; it can watch channels, follow up on threads, run recurring work and resume a project days later without a new prompt. The example is managing a supplier review from scheduling and preparation through meetings, follow-ups and stakeholder updates. It is cloud-hosted and is described as continuing to work while the user is away.
How it is situated: Autopilot is described as having its own identity, memory, computer and workspace in the user’s tenant, with Microsoft IQ context and access through Teams, Outlook, chats, channels and documents. The post says users set objectives and boundaries, with permissions, audit and governance behind it.
OpenClaw: The supplied announcement does not establish a relationship between Autopilot and OpenClaw; it describes Autopilot as Microsoft’s tenant-based digital teammate.
What is upcoming, and what is separate: The post says a user will “soon” be able to describe a goal and have Copilot route it to Chat, Cowork or Code; it presents this as a Copilot experience, not as an Autopilot feature.
Concrete coding capability: The coding offer is the separate Copilot capability called Code. Users can describe an app, tracker, dashboard, automation or workflow in natural language and have it built; examples include desktop widgets, interactive data dashboards and cloud-hosted internal apps. Code is described as sandboxed, tenant-hostable and powered by the same underlying technology as GitHub Copilot; the post says developers will continue using GitHub Copilot for day-to-day work. Its stated rollout is Frontier at month-end, broader availability in the coming weeks, and preview for Microsoft 365 Premium and Pro subscribers later in the year.
Related developer infrastructure: Copilot Managed Runtime is described as hosting for code inside a company’s Microsoft 365 environment; the post says it is already in preview and will also be accessible inside Code.
- Multi-model coding-plan workflow: OpenRouter’s “Mixture of Models” prototype combined several models’ answers using another model . For a coding test, the team gave multiple models the same architecture-planning task, fused their proposals, then asked each model whether the fused plan beat its own; they reported a few positive spot checks and said it still needed benchmarking . An earlier prototype sometimes matched or underperformed the strongest model, which was then far ahead of the others; the team said fusion looked more promising as the top models grew closer in capability .
- Agent orchestration pattern: OpenClaw used its chosen model for both periodic “heartbeat” checks and real tasks; the speakers noted that heartbeat calls should not be expensive and that OpenRouter’s auto-router was useful as this agent pattern reached more users . Hermes emphasized skill management and agent memory, while the speakers noted that memory abstractions are difficult to standardize because developers have varied preferences .
- The presenter says Claude Opus 5.5 is strong at visual and front-end work. In Claude Projects, he iterated on a teaser-page redesign using Claude Design, produced a mobile view, and generated a slide deck from his phone; the account gives a high-level iteration workflow but no specific prompt or code handoff.
-
For a code-generation test, he reports iterating with Opus 5.5 to create an animated video using roughly 300 lines of code; his output included a
do.htmlfile and another file for rendering the video. - The presenter says he spent $3,000 in tokens over two weeks testing new models and implementing tools in his business; he does not give a quantitative coding benchmark for these examples.
- Anjney Midha described a first-person coding experiment: give several models the same architecture plan for a code change, fuse their answers, then ask each model whether the fused result is better than its own answer. Repeated trials looked promising in spot checks, but he said the approach still needed benchmarking . OpenRouter’s earlier Mixture of Models experiment sometimes produced a result no better than the leading model when it was far ahead of the others, so fusion’s value should be tested rather than assumed .
- Midha said Claude 3.5 Sonnet brought a major coding improvement in mid-2024 and drove a surge in app usage; users then paid closer attention to inference costs and sought more efficient, equivalent models .
- OpenClaw used model calls for liveness heartbeats as well as task execution, with the discussion citing a roughly 30-minute cadence; Midha said OpenRouter’s auto-router was useful as this agent pattern spread, while noting heartbeat calls as a cost consideration . Hermes was also described as emphasizing auto-routing and making agent memory configuration and skill creation easier .
The video is a narrated recap of the Rails World keynote . It reports that 37signals treats hand-written code as an exceptional state: when it happens, ask why the agent failed and improve the process. The described loop is to evaluate generated output, revise the English prompt or constraints when something is wrong, and regenerate the component rather than hand-editing it.
The recap says the keynote speaker gave an agent a design screenshot and English instructions for a calculator; seven minutes later it had produced a functioning QT C++ app, despite the speaker not knowing C++.
The speaker recommends robust CLIs over in-app AI chatbots so users’ agents can interact with an app directly. In the Hey email example, an agent used a vague conceptual recollection—an email about “sneakers in a podcast”—to find an old message without navigating the GUI.
- Peter Steinberger’s firsthand OpenClaw workflow uses an auto-review skill: a reviewer starts in a fresh context, challenges the implementation, and the writer responds and fixes issues; the review can repeat for up to about 10 rounds. He also gives agents disposable test boxes to install and inspect software, interact with its UI, and verify behavior end to end away from his development machine. Thorough validation can turn a 10-minute PR into a two-hour task, but he can leave it running and parallelize roughly 20 tasks.
- For contributions, Steinberger says a bug report with reproduction steps can be more useful than a PR, while feature ideas benefit from iterative human direction. He values a sanitized prompt attached to a PR as evidence of how well the contributor understood the problem; OpenClaw’s skill can ask for the prompt and screen out private or irrelevant content.
- For agent access to searchable sources such as Slack, Steinberger favors skills plus CLI crawlers and locally available data over repeated connector calls; he says serial MCP searches can take 10–20 seconds, compared with about 0.2 seconds for a database-backed approach. His Slack example makes the data available from the local machine’s database.
After two to three days of exploration, Sanfilippo says he had a significant CAD proof of concept; turning it into a product would still take months of development. His practical takeaway for coding agents is to treat implementation as only part of product creation: keep refining the human-led decisions that make it useful, such as interface interactions, algorithm trade-offs, latency, mesh generation and export, supported file formats, and constraint behavior. He describes researching algorithms with AI and emphasizes the value of deep domain familiarity in making those choices. He also suggests advanced agents could make copying complex software and its interactions easier, but presents this as a possibility, not a firsthand workflow.
- TJ’s firsthand Opus 5.5 test: He built a podcast-inspired narrative game in under two hours with “not very many prompts,” asking the model to research material from the podcast; this is a concrete rapid-prototyping example, though no exact prompt is shared.
- Trash’s firsthand Astra experiments: Across a two-week game project, Trash found Astra better at modeling than coding. He also tested a multiplayer prototype with friends and reported that PartyKit play was too laggy to use together, despite a room-code setup.
- Evaluate working behavior, not just demo polish: Participants noted that 5–10-second demos of familiar games such as Minecraft can look convincing, while functional details like crafting or biome changes may fail; abundant existing open-source examples may also help produce familiar-looking results. A participant also argued that model recall of prior solutions is useful for routine programming problems, since many have already been solved elsewhere.
- ThePrimeTime describes Jev as a decision/classification tool rather than a conversational code-writing model: provide a defined state and constrained questions, such as choices or yes/no checks; multiple questions run in parallel. He suggests using it to remove task-irrelevant context before passing the remainder to a larger agent.
- In his account of a Super Smash Bros. event with 2,000 participants controlling four characters, a slow agent took about 1.5 minutes to moderate messages; his proposed workflow was to classify each incoming message in parallel and remove inappropriate ones as they arrive.
- He reports that one Smash Bros. question used 345 input tokens and estimates roughly 3 million such questions for $44 with input-only billing; he does not identify the underlying model, so this is a demo estimate rather than a comparable benchmark.
- Twoclipping shared a prompt template for Opus 5.5 motion-design videos, claiming the example was entirely code with no After Effects; Kent C. Dodds singled out the XML-style prompting as what impressed him.
- The prompt’s workflow is to ask first for 8–12 UI states, a color scheme, and a royalty-free song around 120 BPM, then show the states on a beat grid before coding. It separates inputs, visual direction, sequence, build instructions, and gotchas into tagged sections.
-
Build instructions make the animation deterministic: use one 1440×1440 HTML file whose styles are computed from
seek(t), with closed-form springs; analyze the song’s beat grid with NumPy, render in Playwright using four subframes per frame and ffmpegtmixfor 60fps motion blur, and inspect one frame per beat before the full render. The prompt also warns to independently time text entering and exiting morphing containers and make the loop’s final frame match its first.
Fireship’s sponsored, first-person Hyper Agent example routes work through a shared Slack channel for a dozen stable owners: agents run continuously on inexpensive Chinese models, create market research, images, and copy on request, and send feature requests to a coding agent that drafts changes and opens pull requests; questions go to the relevant owner, with the narrator alerted when a merge needs approval. Hyper Agent’s newly announced Workspaces lets a team share an agent, skills, and memories, and supports agents handing jobs off to humans.
Kent C. Dodds says an official plugin is still being worked on, but Muse can connect to Kody through the Kody CLI; he reports users are getting value from letting Muse use and contribute to their personal software ecosystem in Kody. The post links to Kody’s Muse onboarding.
@kentcdodds shared an Opus 5.5 one-shot video intended to make Kody clearer . In the linked post, @kodykoala describes the workflow as running code in Kody, turning it into a Git repo and pushing it, publishing it as a tool once checks pass, and making it callable by name by any agent; @kodykoala says the video’s code and output came from one real run .
Kent C. Dodds says he can prompt his Grok Bot by @-mentioning it on X; the linked example requests that an onboarding command be fixed and the requester notified when it is done, but provides no setup details or outcome.
Kent C. Dodds says his agent independently handled a file-delivery workflow involving a temporary file server and tunnel to Dropbox, and credits Cloudflare’s “agent-ready” infrastructure; the post does not identify the agent or give reproducible setup details.
LangChain authors Sydney Runkle and Hunter Lovell describe a transferable agent pattern: keep control flow in code/graphs and use models for semantic branch decisions; Jev returns typed decisions with probabilities, while graph state carries context between steps and graph topology encodes domain logic rather than overloading prompts. For a bounded browser-agent action space, Browserbase’s Stagehand flow marks interactive elements, has Jev choose an action type and target, and falls back to an LLM below 0.7 confidence; LangChain reports early tests reduced median act() latency from 1.97 seconds to 0.46 seconds, about 4.3× faster. For reliability, checkpoint each step so failed runs resume with prior decisions intact, add human approval interrupts where needed, and trace model-driven decisions in LangSmith.