We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
Simon Willison dropped a high-leverage pattern for agent-heavy codebases: have the coding agent generate interactive/animated explanations of how code works to pay down “cognitive debt” (the “black box” feeling you get when agent-written internals stop being intuitively understandable) . His concrete loop: generate a linear walkthrough, then reuse that walkthrough as context to ask for an animation—ending in a playable demo you can inspect and tweak .
🛠️ TOOLS & MODELS
Claude Code — Remote Control now for Pro:
/remote-controlis now available to all Pro users. The intent: start sessions locally in terminal and continue from your phone without breaking flow .Model routing in a real tmux setup (DHH)
- Layout: OpenCode + Kimi K2.5 (via Fireworks AI) on top, Claude Code (danger mode) on bottom .
- Personal router: start “almost all agent tasks” with Kimi for speed, then ask Claude for second opinion / harder work .
-
Omarchy 3.4 launcher:
tdl c cx(Tmux Developer Layout + OpenCode + Claude Code) .
Codex vs Claude (early practitioner signal)
- Uncle Bob Martin: “Codex is definitely faster and probably smarter than Claude” (initial use) .
- Tibo Sottiaux: “Codex is now starting to be associated to speed” .
Verification-first framing (Addy Osmani): argues the “unsolved problem isn’t generation but verification,” making engineering judgment the highest-leverage skill . Also frames the next step as moving from writing code to orchestrating systems that write code (“building the factory”) .
💡 WORKFLOWS & TRICKS
Turn walkthroughs into animations (Willison’s loop)
- Have the agent produce a linear walkthrough of an unfamiliar codebase .
- Paste that walkthrough into a new agent session and request an animated explanation of the hard-to-intuit part .
- Use the result as an explorable artifact (his example shows spiral placement + collision checks for each word) .
Prompt formatting that reliably improves agent output: use checklists (ThePrimeagen)
- “Hey review every file and tell me …” → “always sucks” .
- Rewrite as a checklist (e.g., “review every file and gather context” then “tell me about …”) because “llms LOVE checklists” .
Put stable intent above fast-changing implementation (swyx + replies)
- swyx: prompt engineering is evolving toward “Specification Engineering”—encoding intents/goals/principles as agents get more autonomous .
- Reply synthesis: separate what you want (task) from how (models/tools/strategies that keep changing) .
Write-code-is-cheap ⇒ testing/QA becomes the choke point (Theo)
- Theo’s claim: “Lines of code effectively are free now… Tests matter” .
- He describes a feature pipeline where you can now skip from “user problem” straight to code via an agent (e.g., screenshot → Claude Code → fix), destroying the old funnel—but leaving review, testing, and release as the real constraints .
Agent-run “company OS” pattern (Pulsia)
- Product claim: Pulsia is “an AI that builds and runs companies autonomously,” covering product coding, marketing, emails, Meta ads, and competitive research .
- Nightly loop: a “CEO” instance decides which task to do, executes, and emails a morning summary + next plan; users steer via email/dashboard .
- Scale signals: “91k human messages” and users averaging “15 messages per day” .
- Infra note: founder uses Neon because it’s pay-as-you-go and “very agent friendly” for spinning up and killing databases .
👤 PEOPLE TO WATCH
- Simon Willison — keeps turning agent usage into durable patterns, now with “interactive explanations” as an antidote to cognitive debt .
- DHH — valuable for operator-grade setups (tmux + two-agent stack + model routing + exact launcher command) .
- Addy Osmani — consistently sharp about where the work is shifting: verification/judgment and “factory model” orchestration .
- Theo (t3.gg) — one of the clearest (and most polarizing) narrators of the “code is cheap; shipping isn’t” transition .
- Miguel Grinberg — governance/attribution reality check: CPython has commits co-authored by the
claudeGitHub user, implying LLM usage is allowed (explicitly or via lack of prohibition) .
🎬 WATCH & LISTEN
1) Theo (t3.gg) — “lines of code are free; tests matter; the pipeline is destroyed” (≈21:20–25:31)
A crisp articulation of why agent coding compresses everything before code-writing—and why review/QA becomes the bottleneck.
📊 PROJECTS & REPOS
- Interactive explanations (guide chapter) — Willison’s write-up + example-driven pattern : https://simonwillison.net/guides/agentic-engineering-patterns/interactive-explanations/
- Rust wordcloud research repo + artifacts (Willison)
- Code/report: https://github.com/simonw/research/tree/main/rust-wordcloud
- Walkthrough artifact: https://github.com/simonw/research/blob/main/rust-wordcloud/walkthrough.md
- Interactive demo result: https://tools.simonwillison.net/animated-word-cloud
- Factory model (long-form) — Osmani’s write-up : https://addyosmani.com/blog/factory-model/
Editorial take: Today’s through-line: generation is abundant—the winning workflows convert that into shipped value by upgrading specs, explanation artifacts, and verification loops.
New Agent Engineering essay & talk by @swyx (@aiDotEngineer): Covers why all-in on agents, agent definitions (via @simonw), Six Elements of Agent Engineering, ChatGPT's path to 1B MAU via agents, and why now. Live on @latentspacepod. https://latent.space/p/agent
@swyx: Prompt Engineering evolves to Specification Engineering—encoding intents/goals/principles for autonomous agents, critical as LLMs handle more code generation. Analogies: US Constitution, OpenAI model spec (used for GPT-4o glazegate). Underrated: legal concepts for establishing/modifying/enforcing intents .
Timeless pattern (DSPy): Separate what you want (task) from how (models/tools/strategies) .
Firsthand from @swyx (@aiDotEngineer lead, @latentspacepod).
Pulsia is an autonomous AI agent platform for building/running companies, handling product coding (50% of tasks), marketing, emails, Meta ads, research .
Workflow: Nightly CEO agent (Claude Opus) assesses bugs/customers/metrics, prioritizes/executes tasks via specialized sub-agents (engineering/marketing/etc.), sends daily email summary/plan; users guide via email/dashboard (avg 15 msgs/day/user, 91k total msgs across 1k+ companies) .
Coding examples (live dashboard): Building sales data recognition engine, adding auth, wiring booking page, email drip w/ analytics .
Onboarding: Input idea/existing biz → auto-research → mission doc (as system prompt/memory) → market report → provisions GitHub/Render/Neon DB → builds landing page/internal tools → auto-schedules tasks (editable/deletable via chat) .
Production use by founder (solo, started Nov 2024, $1M ARR): AI agents find/fix bugs, build features from tickets (agents converse), handle investor responses/marketing; prefers agents over hiring .
Infra for agents: Neon DB (pay-per-use, agent-friendly vs Render) .
Access: pulsia.com .
Firsthand from Ben B., Pulsia founder using in production at scale .
Practitioner insights on AI-assisted coding:
- Developer Trash (speaker D, working on legacy Netflix NRDP codebase) embraces vibe coding for iOS apps as side project; plans to use autonomous agents (context: OpenClaw-like) for remote prompting while "chilling, eating pancakes" .
- Trash uses AI to dissect complex legacy codebase, prompting e.g. "what does this do?" despite initial doubts ("AI cannot help me in that code base") and uncertainty on accuracy .
- TJ (speaker C) on Arch Linux: prompted AI to "install all the packages you need to make [virtual cam with OBS] work" – succeeded first try, unaware of side effects .
- Bash Bunny (speaker B) vibe-coded Nightbot schedule command for timezone handling; prefers trad coding for learning/personal blog project to embrace "the struggle" .
99 is an AI integration into Neovim for code generation and tutorials .
Firsthand workflow (ThePrimeagen building/using it in production-like dev):
- Prompt for specific code tutorial: "Hey, I don’t know how to capture video and audio frames... from when in the browser. Give me a breakdown... Please use typescript." → Receives step-by-step TS example with getUserMedia, canvas pixels, audio samples .
- Persists AI responses/tutorials to disk for session recovery (current refactor goal: save/read state including prompts, history) .
'Work' feature: Describe task (e.g., "refactor the in flight request system..."), analyzes unpushed commits, suggests remaining steps/bugs .
AI-assisted testing/review: Generates tests ("write me a test to prove deserialize works") ; reviews changes, flags bugs (e.g., vim.json.decode misuse) .
Refactors with tests passing (zero-test-fail refactor) .
Factory model for scaling agent-based code generation: orchestrate fleets of agents like a production line using clear specs as blueprints, TDD for quality control, and strong architecture to amplify leverage — shared by @addyosmani, Director @GoogleCloud AI .
Next abstraction shift: from writing code to orchestrating systems that write code ("building the factory" for your code) .
Unsolved challenge: code generation/verification, where engineering judgment is highest-leverage skill .
Longer thoughts: https://addyosmani.com/blog/factory-model/ (framing via Grady Booch: “The entire history of software engineering is one of rising levels of abstraction” ).
Theo (t3.gg full-stack TS expert, 15-person team lead) shares firsthand vibe coding workflows using AI agents for production/side projects:
Built Lon (Frame IO alternative, ~16k TS LOC, open-sourced), part-time in 2 weeks without writing code: described API structures/logic to AI, approved proposals; established data loading patterns (pre-warming, hover subscriptions) in
Agent.mdfor codebase-wide application; fixed misses via targeted prompts e.g., "one of the links in the top nav doesn’t have pre-warming working" .T3 Code (his AI coding GUI/CLI alt, 52k LOC): vibe-coded with Julius in 2-3 weeks; turn-based main branch ownership for direct shipping, outpacing OpenAI's Codex app (20+ team) .
Prompt example for folders: "I want to add a folder style structure... Write up a plan for how we will implement this and what the UX will look like" .
Tools: Claude Code, Cursor, OpenAI Codex/Copilot CLI, Code Rabbit (AI review with memory/feedback learning); screenshot user issues → paste to agent → fixed .
Timeless patterns: Code writing now free/flat (bypasses funnel top); bottlenecks shift to scoping user problems, thorough QA/tests, streamlined review/release (smaller/flatter teams win via low inertia); big teams slowed by approvals .
Contrarian: Lines of code irrelevant; average devs weak at QA/review (now critical); build alternatives to slow incumbents .
Simon Willison advocates interactive explanations to reduce cognitive debt from agent-written code, enabling confident reasoning about black-box implementations .
Firsthand workflow using Claude Code (Claude Opus 4.6):
-
Prompted for Rust word cloud CLI:
Create a Rust app that can create "word cloud" data visualizations given a long input text. - Repo: https://github.com/simonw/research/tree/main/rust-wordcloud; initial prompt: https://github.com/simonw/research/pull/91#issue-4002426963.
- Generated linear walkthrough: https://github.com/simonw/research/blob/main/rust-wordcloud/walkthrough.md.
- Pasted walkthrough into new Claude Code session requesting animated explanation of Archimedean spiral placement .
- Result: Interactive demo https://tools.simonwillison.net/animated-word-cloud; animation shows spiral placement attempts with collision checks .
Timeless pattern: Coding agents produce on-demand animations/interactives to explain their own or others' code .
CPython repository includes 8 commits co-authored by the claude GitHub user from Claude Code (Claude Opus 4.5) over the last 6 months, covering a small code portion .
Workflow: Developers allow Claude Code to commit changes to their local CPython clone, then push to GitHub, carrying the Co-Authored-By: Claude Opus 4.5 tag .
Miguel Grinberg (Python community contributor) observes this indicates CPython implicitly allows LLM use in contributions but criticizes attribution to the tool, arguing developers should take full ownership and it robs community learning opportunities .
Python Dev Guide's Generative AI page provides vague guidance without clear policy on code contributions .
Simon Willison (@simonw), creator of Datasette and co-creator of Django, released a new chapter in his Agentic Engineering Patterns guide: using coding agents to build custom interactive and animated explanations to combat cognitive debt.
Link: https://simonwillison.net/guides/agentic-engineering-patterns/interactive-explanations/
This introduces a timeless pattern for agentic engineering to reduce cognitive overhead in codebases.
@ThePrimeagen observes that vague prompts to opencode like "Hey review every file and tell me …" "always sucks" , but reformatting into checklists—e.g., "[ ] - review every file and gather context [ ] - tell me about …" —works great since "llms LOVE checklists" .
Timeless tip: Use checklists for LLM-based code review tasks .
Claude Code's Remote Control feature, invoked via /remote-control, is now available to all Pro users (previously Max research preview) .
Enables starting local sessions from the terminal and continuing them from your phone without losing flow .
Announced by @_catwu (Claude Code + coworker @anthropicai), quoting @noahzweben .
DHH (creator of Ruby on Rails & Omarchy, CTO @37signals) shares his tmux dev layout for coding agents: opencode (Kimi K2.5 on Fireworks AI) on top, Claude Code (danger mode) on bottom .
Workflow: Starts almost all agent tasks with Kimi (so fast!), then Claude for second opinion/more advanced work .
Launch command with Omarchy 3.4: tdl c cx (Tmux Developer Layout + OpenCode + Claude Code) .
Firsthand production workflow from top practitioner.
Uncle Bob Martin (@unclebobmartin), renowned software engineer, states that Codex is definitely faster and probably smarter than Claude, based on his initial use: "Codex is definitely faster and probably smarter than Claude. I’m pleased with it so far."
@thsottiaux, likely on the Codex team at OpenAI, notes Codex is now associated with speed and anticipates rapid improvements: "Excited for the team to keep building here" .