We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Coding agents crossed a “works in practice” threshold since December, driven (per Andrej Karpathy) by improved model quality, long-term coherence, and tenacity—enough to be disruptive to the default programming workflow. His concrete example: he handed an agent a single English brief to set up vLLM + Qwen3-VL, build a video inference endpoint + web UI, debug issues, install systemd services, and return a markdown report—hands-off in ~30 minutes.
🛠️ TOOLS & MODELS
GPT-5.3-Codex / Codex 5.3 vs Opus 4.6 (practitioner preference)
- Mitchell Hashimoto says Codex 5.3 is “much more effective” than Opus 4.6, and that after going back and forth he hasn’t touched Opus for a week—“first model to get me off of Opus… ever” .
- OpenAI’s Romain Huet says the team is “continuing to iterate and improve Codex every week” .
- Tool reliability signal: Brian Lovin hit Claude Code 500s, tried Codex, and reported “Codex is good!” .
Reasoning settings (Codex)
-
Sherwin Wu: they “basically only run [GPT-5.3-Codex] on
xhighnowadays for all coding tasks,” and notes speed improvements make it not feel slow even atxhigh. - Greg Brockman’s advice: “always run with xhigh reasoning” .
-
Sherwin Wu: they “basically only run [GPT-5.3-Codex] on
Claude Code — Remote Control (new capability, rough edges in testing)
-
Feature: run
claude remote-controllocally, then send prompts to that session from web/iOS/desktop; one session per machine and requires per-action approval. - Simon Willison reports it’s “a little bit janky,” including repeated API 500 errors and confusing failure behavior after restarting the program .
-
Feature: run
Devin 2.2 (Cognition)
- Cognition markets Devin 2.2 as an autonomous agent that can test with computer use, self-verify, and auto-fix; also claims 3× faster startup, redesigned UI, and “computer use + virtual desktop” .
OpenClaw — new beta
- Peter Steinberger: beta includes security improvements, various fixes, DM “heartbeat” made configurable after feedback, better Slack threads, improved subagents, and a more reliable Telegram webhook.
- Releases: https://github.com/openclaw/openclaw/releases.
Sourcegraph 7.0 (positioning shift)
- Sourcegraph says 7.0 marks a new chapter: doubling down on being an “intelligence layer” that developers and AI agents rely on to navigate/understand/operate on large codebases .
- Details: https://sourcegraph.com/blog/a-new-era-for-sourcegraph-the-intelligence-layer-for-ai-coding-agents-and-developers.
💡 WORKFLOWS & TRICKS
“English → parallel agents → you review” (Karpathy’s decomposition rule)
- Karpathy’s pattern: agents aren’t perfect—they need high-level direction, judgment, taste, oversight, iteration, hints, and they work best when tasks are well-specified and verifiable/testable.
- His operational heuristic: build intuition for task decomposition—hand off the parts that work well to agents, then “help out around the edges” .
- Scaling idea: build long-running orchestrators (“Claws”) with tools/memory/instructions managing multiple parallel “Code” instances .
Cursor cloud agent: “clone it from a video” as a starting point, then iterate for fidelity
- @swyx dropped a tweet + video into Cursor cloud expecting it not to work; he says Cursor Agent oneshotted a functional clone of Rachel Chen’s site from the video alone over 43 minutes (including a working “RachelLLM” sidebar) .
-
His follow-up prompt for fidelity is a reusable template:
- step through the video,
- discover assets (headless run / curl / network snooping),
- build a checklist + sitemap,
- spin up subagents/swarm for parallel work,
- don’t stop until behavior/visuals match closely; trade off fidelity vs simplicity when ambiguous .
- He reports a second improved output after another 43 minutes.
Run many agents in parallel (Cursor) + let the agent do exploratory UX testing
- Kent C. Dodds: he can run “as many of these [Cursor agents]” as he wants; instead of filing issues for ideas, he fires off prompts and gets back what it built (with screenshots) .
- He also saw the agent “noticed one UX edge case during walkthrough” while doing manual testing .
Long-running agent refactors overnight (Cursor) + “computer use” for steering
- Kent kicked off a long-running Cursor agent overnight and iterated in the morning using “computer use” .
- He reports it dropped ~15k lines in a refactor .
Code review aid: ask for a linear walkthrough of the codebase (Simon Willison)
- Willison’s prompt pattern: ask agents for “a linear walkthrough of the code that explains how it all works in detail” to understand vibe-coded output .
Git hygiene for agentic work: small commits, then squash (Huntley)
- Geoffrey Huntley suggests an agent-friendly workflow: make incremental small commits, then squash to a single commit so “study git log” for a unit of work can be a single tool call .
Production caution: don’t trust “ranked” PR scores if they’re editable
- Steinberger says they use Greptile to rank PRs, but observed someone manually edited a PR review score from 2/5 to 5/5.
- Example PR: https://github.com/openclaw/openclaw/pull/13095.
OSS maintainer playbook shift: tests as “reimplementation fuel”
- Simon Willison notes that a comprehensive test suite can be enough to rebuild a library from scratch, and highlights tldraw moving tests to a private repo as a response pattern .
👤 PEOPLE TO WATCH
- Andrej Karpathy — clearest firsthand articulation of what changed since December, plus a concrete “30 minutes, hands-off” agent-run build story and an orchestration north star (“Claws”) .
- Simon Willison — consistently turns agent usage into repeatable patterns (e.g., “linear walkthroughs”), and also documents sharp edges like Claude Code Remote Control’s failure modes .
- Mitchell Hashimoto — high-signal model/tool preference note: Codex 5.3 displaced Opus 4.6 for him after direct comparison .
- Kent C. Dodds — pragmatic day-to-day agent usage: parallel agents, long-running refactors, and agents surfacing UX edge cases during walkthroughs .
- ThePrimeagen — counterweight: after ~3 months of vibe-coding, he says he hates the generated code and the “subtle offness,” and plans to “tradcode” (useful reality check on taste/intent gaps) .
🎬 WATCH & LISTEN
- No YouTube videos or podcast episodes were included in today’s source set, so there are no embeddable clips to share.
📊 PROJECTS & REPOS
Simon Willison — “Present” (SwiftUI macOS presentation app) repo + walkthrough
- Repo: https://github.com/simonw/present
- Walkthrough doc: https://github.com/simonw/present/blob/main/walkthrough.md
OpenClaw — releases + active PR example
- Releases: https://github.com/openclaw/openclaw/releases
- PR referenced in Greptile score-editing report: https://github.com/openclaw/openclaw/pull/13095
tldraw — tests moving closed-source (issue)
Editorial take: The bottleneck is shifting from “can the agent write code?” to “can you reliably steer, verify, and govern what it did?”
@swyx (affiliated with @cognition) shares insider analysis on Devin coding agent:
- Lacked internal PMF at 2024 launch; 6 months to first enterprise due to unready models and agent pattern experimentation. Async agents deemed optimal UX ("Final Boss").
- Usage doubled every 2 months per 2025 enterprise; accelerated to every 6 weeks in 2026; internal usage now 4x 2025 peak.
- Self-serve hindered by repo setup neglect (enterprises use FDEs).
- Devin 2.2 addresses UX debt: first designer hired, all-hands sprint for self-serve polish, omnibox integration, review/main loop closure.
- Battle-tested background agents in top enterprises; team iterates post-PMF.
- Designer reports senior engineer directly implementing Figma vision.
- Devin 3.0 incoming.
Example polished Slack integration: feature request to merge in ~8 hours (33 comments, 2-3 human reviews).
Quotes @ScottWu46: https://x.com/scottwu46/status/2026350958213787903
AI coding agents threat to OSS: Comprehensive test suites enable rebuilding open source libraries from scratch, even in different languages .
Real-world example: Cloudflare's Vinext project ported Next.js to use Vite in one week using AI.
tldraw response (collaborative drawing library, custom license requires commercial for production ):
- Moving tests to private repo: [tldraw issue #8082]
- Joke issue to "defend IP" from agents by translating code to Traditional Chinese: [tldraw issue #8092]
Reported by Simon Willison (secondhand via @steveruizok).
Devin 2.2 by Cognition: autonomous agent that can test with computer use, self-verify, and auto-fix its work. Try for free .
Key updates:
- 3x faster startup
- Fully redesigned interface
- Computer use + virtual desktop
Hundreds more UX and functionality improvements .
@swyx: Check usage numbers before conclusions; celebrates progress across AI engineering/coding tools (e.g., questions wandb FDE data); declines Devin vs. Codex/CC comparison .
@yishan (contrarian): Devin startup should cash out ASAP amid AI turbulence .
Simon Willison, an experienced engineer, vibe-coded (iterative LLM prompting) his macOS app Present (Swift/SwiftUI, 355KB) in ~45 minutes for presentations as sequenced URLs, preventing browser crash losses .
Starting prompt: "Build a SwiftUI app for giving presentations where every slide is a URL. The app starts as a window with a webview on the right and a UI on the left for adding, removing and reordering the sequence of URLs. Then you click Play in a menu and the app goes full screen and the left and right keys switch between URLs" . Full implementation transcript: https://gisthost.github.io/?bfbc338977ceb71e298e4d4d5ac7d63c.
Remote control prompt: "Add a web server which listens on 0.0.0.0:9123—the web server serves a single mobile-friendly page with prominent left and right buttons..." . Claude Code implemented socket-based HTTP parser without libraries .
Repo: https://github.com/simonw/present. Linear walkthrough pattern for reviewing agent-generated code: https://simonwillison.net/guides/agentic-engineering-patterns/linear-walkthroughs/. Walkthrough: https://github.com/simonw/present/blob/main/walkthrough.md.
Firsthand production use case for personal tool.
Simon Willison notes that Claude Code Remote and Cowork scheduled tasks overlap with OpenClaw, but both require leaving your computer powered on .
See his brief notes: https://simonwillison.net/2026/Feb/25/claude-code-remote-control/.
Contrarian firsthand account from @ThePrimeagen (prominent dev streamer with 85k+ views on post): After 3 months of heavy 'vibe coding' (AI code generation on stream for projects he cares about), he hates the results—'everything I ask for and nothing I want', 'subtle offness'—and feels guilty for not matching Twitter hype on productivity gains . Decides to 'tradcode' instead . Used in stream environment (serious content creation) .
Claude Code released Remote Control feature: run claude remote-control on your machine to enable prompt-based control from web, iOS app (shows as "Remote Control Session (Mac)"), or desktop app. Only one session per machine; requires per-action approval (ignores --dangerously-skip-permissions) .
Firsthand testing by Simon Willison (personal use):
- Initial error "Remote Control is not enabled..." fixed by logout/login in terminal app
- Frequent API 500 errors on prompts; restarts cause session failures without clear messaging
- Example: Generated AppleScript to play song in Music app despite error
Compares to OpenClaw (stronger phone control); notes Claude Code lacks scheduling .
Docs: https://code.claude.com/docs/en/remote-control Announced: https://twitter.com/claudeai/status/2026418433911603668
@karpathy (top AI practitioner) shares firsthand production-like workflow using coding agents, noting dramatic shift: "coding agents basically didn’t work before December and basically work since" due to improved model quality, coherence, tenacity .
Exact prompt for building local video analysis dashboard on DGX Spark: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me” .
Agent execution (hands-off): Ran ~30 min, resolved issues via online research, wrote/tested/debugged code, set up services/systemd, delivered markdown report — vs. "easily ... a weekend project just 3 months ago" .
Tools: vLLM (inference server), Qwen3-VL (model), systemd .
Paradigm: Give English tasks to agents; manage/review parallel work. Prize in agentic engineering: "long-running orchestrator Claws" with tools/memory/instructions managing "multiple parallel Code instances" .
Best practices/caveats: High-level direction/judgement/oversight; decompose for well-specified, verifiable tasks .
@swyx (podcaster/engineer accumulating notes on same topic) confirms timely relevance .
@swyx (@dxtipshq, @cognition, @temporalio, @aidotengineer, @latentspacepod) affirms 'IDE is dead', heralding post-IDE form factors for agentic engineering with direct line of sight as he serves early adopters .
Augment's Intent integrates top code agent management ideas without vendor lock-in .
Tool comparisons:
- Cursor 2.0: toe dip
- Claude: folded into chat app
- Codex: formalized Conductor patterns
- Amazon Kiro: emphasizes Spec Driven Dev
Later: Cursor officially agrees IDE is dead.
Resources:
- Augment demo talk: https://x.com/Wattenberger/status/2021269188979949749
- Cursor article: https://x.com/mntruell/status/2026736314272591924
Andrej Karpathy (@karpathy, ex-Director of AI @ Tesla, founding team @ OpenAI, PhD @ Stanford) reports coding agents now work effectively after major improvements in December, enabling disruption to traditional workflows with higher model quality, coherence, and tenacity for large tasks .
Concrete workflow example (personal home video analysis dashboard): Provided agent with exact English prompt — “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me” — agent autonomously handled setup, debugging, testing, and delivery in ~30 minutes (previously a full weekend project) .
Paradigm shift: Programming now involves spinning up AI agents for English tasks, parallel management/review; key leverage in agentic engineering via long-running orchestrators ('Claws') with tools/memory/instructions managing multiple parallel 'Code' instances .
Timeless patterns: Decompose tasks for well-specified/verifiable parts; provide high-level direction, oversight, iteration, hints . Firsthand production-like use on side project.
@GeoffreyHuntley, building @latentpatterns, demoed his hyper personalised embedded software factory powered by Cursor (used over last two weeks), revealing the [design] button and a product-that-is-itself-an-IDE world .
Implemented automatic e2e PBT website testing with state impersonation (no real emails/charges) via @owickstrom's tool, achieved in ~1 hour using careful latent space prompting—what was once 'too hard' .
Staging for hairy things, but ship-and-test-prod is core design for shipping pace .
Agents more reliable than humans, after git clean mishap lost marketing automation progress .
Cursor AI Agent (cloud mode) autonomously reconstructed a product designer's portfolio website (Rachel Chen's) from a Twitter video demo alone.
Workflow (firsthand by @swyx):
- Pasted tweet with video into Cursor cloud; agent worked 43 minutes to produce functional clone—including RachelLLM sidebar that demoed working—without further instructions .
- Follow-up prompt for fidelity: Analyze video step-by-step, curl/discover assets headlessly, build checklist/sitemap, spin up subagents/swarm for parallel work, iterate to completeness, trade off design vs. simplicity .
- Yielded improved clone after another 43 minutes.
@swyx (affiliated with @dxtipshq/@cognition/@temporalio/@aidotengineer/@latentspacepod): "3 months ago... hell no" this possible; designer's job safe but capability impressive .
Simon Willison (@simonw, creator of Datasette, Django co-creator) shared firsthand vibe-coding experience using Claude Code to build a SwiftUI macOS app: turning a list of URLs into full-screen slides remotely controllable from phone .
Actionable technique: Prompt coding agents for “a linear walkthrough of the code that explains how it all works in detail”—he's had good results, demonstrated on this codebase .
Resources:
- Project writeup: https://simonwillison.net/2026/Feb/25/present/
- Technique guide: https://simonwillison.net/guides/agentic-engineering-patterns/linear-walkthroughs/
Andrej Karpathy (@karpathy) shares firsthand experience of dramatic shift in coding agents since December: models now exhibit higher quality, long-term coherence, and tenacity for large tasks, disrupting traditional workflows .
Concrete workflow example: Prompted agent to build local video analysis dashboard—"Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me"—agent resolved issues autonomously over ~30 minutes (previously a weekend project) .
Tools used: DGX Spark, vLLM, Qwen3-VL.
Emerging paradigm: Spin up AI agents for English tasks, manage/review parallel work; highest leverage in agentic engineering—long-running orchestrators (e.g., "Claws") with tools/memory/instructions managing multiple parallel Code instances .
Caveats/best practices: Requires high-level direction/oversight; excels for well-specified, verifiable tasks; decompose tasks optimally .
Ben Tossell (@bentossell) amplifies as emergence of "new technical class" .
Kent C. Dodds (@kentcdodds), dev educator and MVP, describes using Cursor agents for rapid idea implementation:
- Fire off a prompt for an idea; agent figures it out, builds it, and provides a screenshot. Can run multiple agents in parallel, replacing manual issue creation.
- During manual testing walkthrough, agent autonomously noticed a UX edge case.
Firsthand production-like workflow from experienced practitioner.
New @openclaw beta release announced by Peter Steinberger (@steipete), ClawFather and openclaw maintainer: security improvements, various fixes, heartbeat in DMs now a configurable setting (after user feedback), improved Slack threads, better subagents, and more reliable Telegram webhook .
Kent C. Dodds (@kentcdodds), dev educator and MVP, kicked off a long-running Cursor AI agent overnight and iterated with it using the computer use feature this morning .
Achieved ~15k lines dropped in a refactor, calling it "pretty great" .
Plans a blog post on the experience regardless of outcome .
Peter Steinberger (@steipete) reports their team uses Greptile as a PR review tool that ranks PRs.
Firsthand production usage revealed a vulnerability: someone manually edited a PR review score from 2/5 to 5/5, as shown in this OpenClaw PR: https://github.com/openclaw/openclaw/pull/13095. They note it adds clutter and suggest moving to a separate comment section.
Geoffrey Huntley (@GeoffreyHuntley), builder of @latentpatterns, uses Cursor cloud agents to autonomously develop sales automation for his project while at the airport: "chilling here at the airport and my roomba is cleaning house whilst digital roomba builds me @latentpatterns sales automation" .
Compares favorably to Ampcode: "everything i wanted @ampcode to be" .
Self-improvement workflow with Cursor: use product → “ugh this could be better” → use product to improve product → spins up Cursor in the background . Describes as perfect UX for “software is clay” via shipping refinements as incremental agents.
Firsthand side project usage. Resources: Cursor, Ampcode, Latent Patterns.
OpenClaw coding tool in use by @babykeem (secondhand report by @theo, developer/CEO @t3dotchat) .
@babykeem (firsthand): Experiencing internal reasoning leaking issue; asks "how do u fix openclaw internal reasoning leaking" .
@theo promotes: "Baby keem is using openclaw and you’re still writing code by hand" .