When Claude Deleted My Tests: Jesse Vincent, Creator of Superpowers

Software used to cost something to write. Now the hardest part is knowing what to do once it doesn't. Harper Reed (https://www.linkedin.com/in/harperreed), CEO at 2389 Research (https://www.linkedin.com/company/2389-research/), joins us to talk about what happens to software development when the cost of writing code drops to zero. Harper built Test Kitchen, a tool that runs the same feature five different ways and lets you judge the winner, and he's candid about the tools he's built in the past that scaled in directions he didn't expect. KEY TAKEAWAYS 00:00 Harper Reed's biography spans juggling, anarchism, the Obama campaign, and a social network built for AI agents 01:06 Harper built his AI agents and their own social media platform before this became an industry trend 12:21 Early evals found tasks were completed more effectively with the social media plugin installed than without it 16:40 The cost of writing code is approaching zero, and that changes everything about process 17:33 Test Kitchen builds the same feature multiple times and lets you or an AI judge pick the winner 23:16 Harper defines exciting software as something with personality, not raw speed 24:15 Harper built tools that later fed directly into the Cambridge Analytica playbook 31:10 Harper argues the industry does not ask often enough whether it should build something 49:46 Omakase means trusting the model’s judgment instead of specifying every decision yourself 50:10 2389 Research has built roughly 50 internal skills, most centered on parallel exploration and judged review 52:04 2389 Research was founded on a bet that AI agents can make work more human, not less See how people are actually using AI in their work on the "AI Worth Using" podcast: https://aiworthusing.com. RESOURCES MENTIONED 2389 Research website: Harper Reed's company, where the tools and experiments discussed in this episode were built: https://2389.ai/ Test Kitchen (skill directory): Harper's skill that builds a feature multiple times and uses an AI judge to pick the best version: https://2389.ai/products/test-kitchen/ Omakase (Test Kitchen's parallel-build workflow): Test Kitchen's parallel-exploration mode: hands a model a spec and lets it build with minimal guardrails: https://2389.ai/posts/omakase-show-me/ Binary RE (binary reverse-engineering skill): A binary reverse-engineering skill Harper says is strong enough that Fable refuses to run it: https://github.com/2389-research/binary-re/blob/main/skills/SKILL.md Thrifty (heavier-model planning, lighter-model execution skill): Uses a heavier model to plan and a lighter model to execute, cutting token cost by about 60%: https://github.com/2389-research/thrifty Deliberation (Quaker-practice-inspired agent skill): Built by co-founder Dylan; applies Quaker consensus practices to how agents make decisions: https://skills.2389.ai/plugins/deliberation/ Review Squad (multi-agent review skill): A lightweight multi-agent review pass run on finished work: https://github.com/wan-huiyan/agent-review-panel 2389 Research paper on agent social-media use and task performance: https://arxiv.org/pdf/2509.13547 #AIagents #BuildInPublic #SoftwareEngineering

Most builders already have more intelligence at their fingertips than they know what to do with. That's the tension Sam Odio and Tom Preston-Werner press on with Peter Steinberger, Creator of OpenClaw, who argues agent security is a trade-off, not a guarantee, no matter who built the model. The creator of OpenClaw would rather have your prompt than your pull request. What he reads in it: whether you understood the problem, or whether you are just a slot machine. Peter also gets into the security trade-offs of ageal difference is marketinentic AI, why open source can match closed labs on protection, and how he keeps dozens of parallel coding agents productive without losing control of the codebase. KEY TAKEAWAYS 00:00 A code reviewer called 600 commits in a day from an AI agent "actually not slop" 01:39 One founder handed a fully-capable agent to strangers on Discord overnight, and it survived 04:32 Once you're a builder, the instinct never fully goes away, no matter how burned out 05:46 The insight behind OpenClaw wasn't new information; it was a feeling many builders shared 10:15 A risk calculation, not bravado, led to putting an unrestricted agent on a public server 13:50 The gap between what AI can do and what we actually use it for keeps widening 20:32 No coding agent, open or closed source, is protected against a determined enough attacker 27:27 Open source security can match closed labs; the rg, not protection 34:14 A fresh-context review agent can catch a coding agent's mistakes before a human sees them 41:11 The best advice for new engineers: stay curious and get involved before mastering the codebase 47:34 Plow Latch gives any agent access to a Mac, gated by one adversarial reviewer 51:40 Tom's first live command to his new agent: say hello world out loud on his Mac 1:00:20 Sam's daily use: paying bills from a photo, untangling scheduling emails, catching a missed mortgage payment 1:06:48 Limited release lands in about a week, with a community launch roughly a month out See how people are actually working with AI at https://aiworthusing.com RESOURCES MENTIONED OpenClaw, the open-source personal AI agent project Peter founded and now stewards through the OpenClaw Foundation: https://openclaw.ai/ OpenAI homepage: https://openai.com/ Nutrient (formerly PSPDFKit), the PDF framework Peter bootstrapped and ran for 13 years before stepping away: https://www.nutrient.io/sdk/ "Where Good Ideas Come From" by Steven Johnson, the book Tom references on the adjacent possible: https://www.betterworldbooks.com/product/detail/where-good-ideas-come-from-the-natural-history-of-innovation-9781594487712?shipto=US&curcode=USD&utm_source=google&utm_medium=cpc&utm_campaign=pMax-Used-USA&utm_content=&utm_term=&creative=&device=c&placement=&gad_source=1&gad_campaignid=16440140764&gbraid=0AAAAADMQnG_QXQP7OowB-WRdNT19pR48_&gclid=Cj0KCQjw5P7UBhDaARIsAOSlS1MaPAsp5uZhTBbt0DFtGwF2xkYRk7ptR0wTLQC1IURuiRDZnWyP8Y4aAtzrEALw_wcB Plow Latch, the Mac-based MCP server demoed live in this episode's bonus segment Download and get started here: http://plow.co/latch Join the AI WorthUsing x OpenClaw 2.0 https://luma.com/zhkhsnpa ​The Hackathon winner, up to 2 members of the winning team, will be flown to SF on October 6th to workshop their agent with AgentCribs, a community of Silicon Valley agentic software builders and present it live to the developer community. Also attended by Peter Steinberger & Tom Preston-Werner. Join the show Discord for more great resources and hackathon with great prizes. https://discord.com/invite/fDY2bBThRs?utm_source=luma #OpenSource #AIAgents #AgenticEngineering

The creator of pandas doesn't read his own code anymore. His review system does, and it trusts nothing.

AI Worth Using is where vetted engineers screen-share the setups they actually run. Live installs, real configs, nothing staged. Samuel Odio and Tom Preston-Werner co-host the AI Worth Using podcast. Live installs, real configs, no hype, zero sponsored takes. Agents now write code faster than we can review it, and everyone has a hot take about what to do next. Almost nobody shows their workflow. So every episode a guest installs their tool on our machine, we learn it on camera, and you leave with something you can copy instead of an opinion to argue with. KEY TAKEAWAYS 00:00 AI agents now write code faster than we can review it — that gap is the whole reason this show exists. 00:08 AI Worth Using cuts through hot takes by showing real, unstaged workflows instead of talking about them. 00:15 Every guest screen-shares their actual setup live — real installs, real configs, nothing staged for the camera. 00:24 Wes McKinney, creator of pandas, runs 150 automated code reviews a day through his own pipeline. 00:24 Wes's adversarial code-review pipeline is fully open-sourced, and he shows the live queue on the show. 00:34 Jesse Vincent's Superpowers is the most-used engineering skill library built for Claude Code. 00:34 Superpowers has been cloned 1.7 million times in just two weeks. 00:34 Hosts Tom Preston-Werner (GitHub co-founder) and Sam Odio (shipped AI at Facebook and Dropbox) bring real builder credibility. 00:48 No sponsor takes and no vaporware — just tools the hosts actually use. 00:55 New episodes of AI Worth Using drop biweekly at aiworthusing.com. RESOURCES MENTIONED Samuel Odio LinkedIn https://www.linkedin.com/in/samodio/ Tom Preston-Werner LinkedIn https://www.linkedin.com/in/mojombo/ Wes McKinney LinkedIn https://www.linkedin.com/in/wesmckinn/ Jesse Vincent LinkedIn https://www.linkedin.com/in/jessevincent/ AI Worth Using — show website: https://www.aiworthusing.com See how people are actually working with AI at https://aiworthusing.com/builder-index #AITools #BuildInPublic #OpenSource