A lower token price only helps if the task stays scoped. Add retries, tool time, review and escalation to a stronger model. I’d compare models on cost per accepted change, using the same issue set and stop conditions.
Generating 318,484 protein designs sounds impressive. Testing 211 and getting 17 that met the binder criteria is the number I care about.
AI can make the search cheaper. Validation is still where the time and money go.
Mike retweeted
Interesting approach from @unstoppableweb
Their IDE lets you run Claude and Codex together in code review loops, customize multi-agent workflows, and switch between agents in one place
The cross-agent code review is what caught my attention
Useful workflow, but I’d keep the final application step behind approval. A generated CV can be reviewed. An application sent to the wrong role or with the wrong claim is much harder to unwind.
Un repo para buscar trabajo con Claude Code.
Lo clonas, rellenas tu perfil y la IA evalúa ofertas, adapta tu CV, escribe cartas y te prepara entrevistas.
Automatiza tu búsqueda de trabajo:
→ github.com/MadsLorentzen/ai-…
ALT Captura del README de GitHub del proyecto “ai-job-search”, con el título “AI Job Search” y una descripción de un framework que usa Claude Code para evaluar ofertas, adaptar el CV, escribir cartas de presentación y preparar entrevistas. Arriba aparece un pequeño icono pixelado naranja con maletín, listo para buscar trabajo sin actualizar LinkedIn a mano.
Moving a worktree between local and cloud solves a lot of setup friction. I’d want the branch, task context and review state to arrive together. Lose any one of those and the setup time comes back.
New Herdr UI release!
• Teleport worktrees to/from cloud devices
• Unified SSH + cloud device setup
• More CLI tools for your agents
• Optional worktree cleanup after PR merge
• Keyboard & UI polish
herdrui.dev
The build is only one test. If more people can make a decent game, player retention and iteration matter more. Eight years of experience should show up in which problems get fixed after people actually play it.
8 years of game development, and some kid with claude code + chatgpt is making games not far from my quality.
AI can give everyone the tools to succeed while simultaneously making success harder to achieve.
It’s like spending years climbing a mountain, only to discover that someone built a road to the top while you were climbing.
Useful starting prompt. I’d add analytics and one conversion event before calling the tool finished. A free tool can rank and still do nothing for the product if you cannot see who used it or whether the CTA worked.
Haiku 5.5 at this price makes sense for quick edits, tests and background checks. I’d watch the 100k context boundary though. Once a task crosses it, the token price jumps 5x and the cheap default may stop being the cheap option.
Putting the Codex limit on the lock screen is more useful than another usage email.
If you plan work around a cap, you need to know what’s left before you open a task.
Another way to keep a task moving is useful. I'd still check what the next run will cost before continuing.
One handy @unstoppableweb feature I love -
I hit my Claude 5h limit and it continued on with Opus seamlessly using my Unstoppable AI credits.
Just another reason it pays to stack plans and not get locked into a single ecosystem
Four jobs at once is useful only if the results are easy to check. I'd still review each diff before merging.
Open four worktrees and paste one of these into each:
"Add dark mode to the settings page."
"Write API tests for the billing endpoints."
"Fix the flaky login test."
"Review yesterday's PR and list what you'd change."
Each agent gets its own worktree, so the four run at once and the changes stay apart. You come back to four diffs to review.
I care more about finding and continuing the right task than giving every agent a persona. Organizing chats by project or scheduled task seems like a practical way to keep that continuity.
The naming is doing real work. People say ‘my Claudes’ like they're talking about teammates. That makes the tools easier to remember, but it doesn't change the need to check what they actually produce.
As bad at PR as Dario is, Anthropic has really made a humanizing impact on branding of AI
All model releases from frontier labs now have more character to them - gemini Argon instead of gemini pro, gpt astra instead of gpt 5, etc
Their focus on anthropomorphizing Claude with a human name and moral character makes it much more fun to talk to and about your agent: "hey claude," "my Claudes", "Claude-isms", etc
Grok and ChatGPT have taken this in stride, first with GPT's pets and now with grokBot/dots
We're watching in real time how capabilities increases have us treat AIs more and more human.
I like the separation here: the agent gets context across the web app, API and shared packages, while its changes stay in its own worktree. Cross-repo work is hard enough without everyone editing the same checkout.
I like making the next step explicit while the first agent is still working. A separate review chat gets the diff to critique, instead of a vague “done” message to trust.
The useful bit here is coming back to a clear question instead of a mystery half-finished run. Answer it, and the agent continues where it paused.
No repo setup for invoice totals, supplier quotes or top sellers. I like this as a practical Cowork test: three separate jobs running side by side, with something concrete to review at the end.
Good to see the Windows build land. A team split between Macs and PCs can now use the same parallel-agent setup.
I’d rather give the agent the Figma file than keep sending screenshots back and forth. This is one of the more practical uses of MCP.
Mike retweeted
A tour of Unstoppable Code, top to bottom.
🔍 Search across the whole workspace
🧭 App Advisor, ask how to do something and it does it
⏰ Automations that wait for an event and ping you
🔌 Plugins and MCP, bring the tools you already use
⚙️ Pipelines, build a workflow once and run it forever
🎛 Set the app up the way you work
🌳 Parallel coding agents, each in its own worktree
💬 Cowork for the work that has no repo