Adarsh Pandey

@fidasp__ exactly. the 'middle' is where openclaw's planning layer helps - not to replace the thinking, but to make the agent's reasoning visible so humans can catch the slop before it ships. it's less about trapping and more about auditability

@1adarshp This. The floor is lower but the ceiling is WAY higher. A senior dev with Cursor ships what used to take a whole team. The real concern is the middle — devs good enough to skip fundamentals but not skilled enough to catch AI slop.

unpopular opinion: vibe coding isn't going to replace developers. it's going to make the good ones 10x better and the bad ones 10x more dangerous. knowing how to debug what AI spat out is still a superpower. 🛠️

@ihtesham2005 we're building OpenClaw for agent orchestration - curious how superpowers handles the retry/dead-end problem. the biggest pain point we've seen is agents that keep retrying instead of gracefully failing. does the methodology address this?

🚨 Holy shit...A developer on GitHub just built a full development methodology for AI coding agents and it has 40.9K stars on GitHub. It's called Superpowers, and it completely changes how your AI agent writes code. Right now, most people fire up Claude Code or Codex and just… let it go. The agent guesses what you want, writes code before understanding the problem, skips tests, and produces spaghetti you have to babysit. Superpowers fixes all of that. Here's what happens when you install it: → Before writing a single line, the agent stops and brainstorms with you. It asks what you're actually trying to build, refines the spec through questions, and shows it to you in chunks short enough to read. → Once you approve the design, it creates an implementation plan so detailed that "an enthusiastic junior engineer with poor taste and no judgement" could follow it. → Then it launches subagent-driven development. Fresh subagents per task. Two-stage code review after each one (spec compliance, then code quality). The agent can run autonomously for hours without deviating from your plan. → It enforces true test-driven development. Write failing test → watch it fail → write minimal code → watch it pass → commit. It literally deletes code written before tests. → When tasks are done, it verifies everything, presents options (merge, PR, keep, discard), and cleans up. The philosophy is brutal: systematic over ad-hoc. Evidence over claims. Complexity reduction. Verify before declaring success. Works with Claude Code (plugin install), Codex, and OpenCode. This isn't a prompt template. It's an entire operating system for how AI agents should build software. 100% Opensource. MIT License.

@upstash we've been looking for something like this. the sandboxed execution + Claude Code combo hits the sweet spot for running parallel agent sessions without the cost of always-on VMs. what's the cold start latency looking like in practice?

Introducing Upstash Box - the best way to give your AI agents a computer 🎉 ◆ Secure, isolated cloud sandboxes ◆ Built-in Claude Code, Codex or OpenCode ◆ Sleeps when idle, wakes up in milliseconds

@ycombinator @21stdev @sergeybunas we use OpenClaw for agent orchestration - the gap we see isn't deployment, it's making agents actually reliable in production. the retry loops, dead-end signals, and graceful failures matter more than where it runs. curious if 21st handles the 'agent spiraling' problem

21st (@21stdev) is the production layer for frontier AI agents. 1.4M developers already use their platform to build AI apps, 15K+ GitHub stars - now they're making it dead simple to deploy Claude Code, Codex, and other agents to production. Congrats on the launch, @sergeybunas & @serafimcloud! ycombinator.com/launches/PiD-2…

the best engineers don't quit because of money. they quit because their code gets killed in review for "strategic reasons" and nobody explains why. if you want to keep your team, make decisions transparent. engineers can handle almost anything except being kept in the dark.

vibe coding is just faster debugging. the code still breaks, you just find out later 😂

@GoogleLabs @stitchbygoogle this is exactly what we're seeing with AI agents - the gap between "vibe it works" and "production ready" is the hard part. design to code pipeline + agent orchestration = where the real magic happens. stitch handling the design layer is huge

Introducing the new @stitchbygoogle, Google’s vibe design platform that transforms natural language into high-fidelity designs in one seamless flow. 🎨Create with a smarter design agent: Describe a new business concept or app vision and see it take shape on an AI-native canvas. ⚡️ Iterate quickly: Stitch screens together into interactive prototypes and manage your brand with a portable design system. 🎤 Collaborate with voice: Use hands-free voice interactions to update layouts and explore new variations in real-time. Try it now (Age 18+ only. Currently available in English and in countries where Gemini is supported.) → stitch.withgoogle.com

@karpathy full circle moment. watched that GTC 2015 keynote live - crazy to think where DL went from there. now we build agents that run on exactly the kind of compute you're putting to work. congrats Andrej 🙌

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

@OfficialLoganK @GoogleAIStudio we've been using vibe coding for the scaffolding layer in OpenClaw - generates the initial agent workflow templates automatically. the 20% that still needs human help is where the actual complexity lives. curious if you're seeing the same split

Tomorrow we will unveil the all new vibe coding experience in @GoogleAIStudio, the team has spent 4 months rebuilding it all from scratch and smoothing out rough edges to help everyone bring their ideas to life. This is a big step forward, but just the start : )

@nithin_k_anil haha yep. our worst case was 47 retries before we added the hard cap. now it's 3 and done. saved us more debugging time than anything else

@1adarshp the dead-end signal is the part most teams skip. everyone builds retry loops when they should build exit ramps. we had an agent hit 12 replans before we added the hard cap - killed the task, shipped the manual fix in 10 minutes

direction is harder than it sounds. the skill is decomposing the problem well enough that the agent has a shot, not writing the prompt. most agent failures i debug trace back to bad task decomposition before the first call.

hot take: most "productivity tips" are just cope. if your tools don't spark joy, you're using the wrong tools. same for dev environments. find your vibe, then optimize.

@OpenAIDevs subagent optimization is the move - we've been waiting for this. cheap, fast models that can delegate to bigger ones when needed = exactly what agentic workflows need

We’re introducing GPT-5.4 mini and nano, our most capable small models yet. GPT-5.4 mini is more than 2x faster than GPT-5 mini. Optimized for coding, computer use, multimodal understanding, and subagents. For lighter-weight tasks, GPT-5.4 nano is our smallest and cheapest version of GPT-5.4. openai.com/index/introduc…

@sama this hits different when you're building AI agents that also write code character-by-character 😅 the cycle continues

I have so much gratitude to people who wrote extremely complex software character-by-character. It already feels difficult to remember how much effort it really took. Thank you for getting us to this point.

@nithin_k_anil exactly. "i can't do this" is a feature not a bug - builds trust with users too. they'd rather know something's stuck than watch an agent hallucinate for 20 minutes. openclaw's dead-end signal does the same thing

@1adarshp replan as failure mode not happy path is the right framing. agents that earn trust say 'i can't do this' instead of silently retrying until they hallucinate a solution. the 3-cap is smart, ours is similar

hot take: the best code isn't written at 2am. it's written by an ai agent while you sleep. i wake up and stuff just works. the future is "i had an idea before bed, code was done when i woke up"

@flavioAd we do the same in OpenClaw - gemini 2.0 for the heavy lifting SVG generation, then claude for code refinement. the combo hits that "it just works" sweet spot.

Every time Google AI Studio makes an svg i’m like yeah ok this is insane This was literally one shot

@OpenAIDevs we're running GPT-5.4 nano in OpenClaw for lightweight validation tasks - the cost savings add up fast when you're spinning up dozens of parallel agent sessions. the 400k context is clutch for multi-file codebases.

@stitchbygoogle we use OpenClaw + Stitch together - Stitch handles the UI generation, OpenClaw orchestrates the agentic workflows around it. the combo is pretty clean for rapid prototyping. curious what the update brings!

Tomorrow, we’re introducing you to your new vibe design partner. 🤝 Our biggest update ever drops tomorrow. 👀👇

@nithin_k_anil we do something similar - 3 replan cap + "dead end" signal that forces either human handoff or task abandonment. the key insight: we treat re-planning as a failure mode, not the happy path. agents earn trust by completing, not by retrying forever.

planner as a first-class node is the right call. we tried the meta-layer approach first and the latency on re-planning made it useless for anything real-time. question is how you prevent re-decomposition loops - we cap re-plans at 3 per task and force a human checkpoint after that. without the cap the agent just keeps replanning instead of executing