... probably saw the comparison: one coding agent harness sends ~33k tokens before your prompt, another sends ~7k. Big thread, lots of outrage about waste. I finally read the whole study instead of the headline, and the actually useful findings are different from what got shared. First, in their realistic-config lane (instruction file + several MCP servers), the "light" harness

... building multi-agent coding workflow or software factories this one is for you! We've been using it internally Kerno for the past 6 months and it has been a GAME CHANGER! We're excited to launch HAR ( https://www.producthunt.com/prod... )! It's an open-source, agent-agnostic framework for building multi-agent coding workflows. Run a fleet of coding agents in parallel on any repository, with deterministic validation gates, verifiable proof, and full observability across every agent, all extensible ...

... exist three years ago. You describe a feature, an agent writes it, and forty minutes later it's running. It works, sort of. It answers, it responds, it does the thing. And you sit there with a genuinely hard question you have no tool for: is it any good? That question used to answer itself. Old software was deterministic same input, same output, and you wrote a test that either passed or failed. AI features don't work like that ... ... same thing twice and you get two different answers, both plausible, one subtly wrong. "It ran without erroring" tells you almost nothing about whether it's right, helpful, or safe. The build got easy. The judging got hard

... this platform. We develop both software and hardware products, but since our software is designed ... ... integrate with the hardware

... Product Hunt Submission Draft ## Product Name remote-browser.dev ## Tagline A remote headed browser for agentic UI testing ## Short Description Build an agentic UI test harness that lets local coding agents drive a visible remote browser, collect screenshots, judge UI expectations, and produce repeatable test reports. ## Article / Maker Comment ### Building an Agentic UI Test Harness

... Codex (now @ChatGPT by OpenAI )is strong at long-context planning and execution. @Claude Code acts like a supervisor and a second pair of eyes. @OpenClaw is like my daily operations intern. I know there are many other agent harnesses. Many are open-source, including OpenHands, @Pi Coding Agent , and Hermes. I m curious. Which agent harnesses

... Almost every knowledge-work domain now has a decent tool. Harvey and Legora for legal. Rogo for finance. All decent and all narrow and not accessible to everyone. Coding has something the rest of knowledge work doesn't: harnesses

... safety, and AI compliance often get used interchangeably. They're related, but I don't think they're describing the same thing. The simplest way I've found to think about it is that AI safety is about reducing harmful behavior. If an AI system gives dangerous advice, produces harmful

... Product Hunt solo builder here. Every multi-agent thing I built ended up with one "manager" object that knew every agent and every step, and an agent harness that was a giant while-loop owning the model call, the tools, the memory, and the approvals. Add a fourth agent and both became the thing that breaks. So I built Cosmonapse , a distributed agent-to-agent protocol where agents are peers on a shared bus. The model is a nervous system ... ... names tell you where logic goes: Neuron thinks (your agent, a pure function), Axon sends, Dendrite reacts, Synapse connects (the bus), Engram remembers. To build harnesses

... standard consumer webcam. No Tobii. No IR sensors. No specialized hardware of any kind.    The immediate question is always: how?    The honest answer is that webcam-based eye tracking is a fundamentally different problem from hardware

... have a different problem how do I scale something? The reasons can vary in nature: legal (taking someone to court), financial (convincing investors), or even existential (where legal costs can put your personal assets at risk). What was the hardest problem you've faced in business so far? Comment from Nika(@busmark_w_nika): @li_yujie  Which channels were good for you? Comment from Anders T.(@atherkildsen): @busmark_w_nika  lol Comment from Anders T.(@atherkildsen): vote farms ... ... channels have become too expensive for startups. Really depends what you're selling and you have to spend a lot of money to find out what does or doesn't work. Comment from Ajin Sunny(@ajinsunny): Distribution is the hardest

... anyone who's felt uneasy trusting a single AI's confident answer on something that actually matters. Hardest part right now: we're pre-launch (Product Hunt, July 14) with close to zero audience, so most of the validation happens one real question at a time instead of through numbers yet. polora.ai Comment from Haritha

... pretty decent. But the real lesson was not how do we get people to sign up? It was: how do we get founders to trust the product enough to actually press send on fundraising emails? That is a much harder problem. For context, Causo helps early-stage founders find relevant investors and run cold fundraising outreach without manually building lists, writing everything from scratch, or duct-taping 5 tools together. Rough numbers so far from the fundraising product: 272 founders ... ... Much fewer were ready to launch outreach. Some were nervous. Some wanted to edit everything forever. Some were probably just curious. Some likely realized they were not ready to raise. That changed how we think about the product. The hard

... fact. we put a control layer in front of every action so risky ones get flagged for human approval before they run, not after. your point about "what's the problem" is exactly right and it's the hardest thing to nail in messaging. demos show capability, but buyers need to see their own pain reflected back at them first. if they don't recognize the problem immediately, the demo doesn't matter. Comment from Volodymyr Zapivakhin(@zapimir): I built ... ... each bad hire costs over 200k. must have problem for HR. three. the 92 percent of work that linkedin never covered. plumbers, nurses, electricians have never had a portable record. that is the most must have segment because the alternative

... Hardware checks. Compatibility scans. Model deployment. Copilot routing. Local hosting. Multi-cloud instances. Cloud failover. Used to take a day. Now under 10 minutes. AI moves fast. Deployment doesn't. 40% of teams take more than a week to get a single model into production. Data scientists spend over a quarter of their working day on setup, not science. That's not an AI problem. That's an infrastructure problem. oneinfer-edge fixes it. Not by reinventing the stack ... ... orchestrating what already exists into one open source control plane. - Multiple serving libraries. One scan. Ollama, llama.cpp, vLLM, SGLang, TensorRT-LLM, PyTorch, Dynamo. Instead of manually testing each one against your model and hardware

... hunters First time posting in p/introduce-yourself. I'm Lisa, co-founder of a cycle tracking app we built in Amsterdam. Wanted to introduce myself and share the harder lessons from the last six months, in case anyone here is wrestling with similar problems. Quick context. In 2021 the FTC fined Flo Health for sharing user period data with Facebook and Google. Millions of women learned, after the fact, that their cycle data when they bled, when they were trying ... ... still has 77 million users. Clue still raised millions. The category just moved on. We couldn't. The question we kept asking was the obvious one: why does cycle tracking have to be ad-funded? So we built the alternative

... until end April this year. Codex mac app + GPT 5.5 was a game changer for me. It magically reduced a lot of back and forth for my coding tasks. Comment from Abijah Kajabika(@abijahkaj): cursor / vs code google chrome gmail slack i'm trying hard

... Product Hunt community! I m Harsha, the founder and systems architect behind Mimic . If you looked at my day-to-day right now, you d see me deep in lines of native Kotlin code, optimizing real-time hardware

... ledger of bets . Downweight-not-delete is the natural operation on a bet. You changed the data model and the hard problem fell out of the frame. The "versioning comes for free" consequence is the part I'd underline, because it's where this beats every hand-rolled approach, including the ones in this thread. Hard

... Feature Update: You can now route your coding copilot or any agentic harness' traffic to self hosted models either in local or cloud in oneinfer-edge (fully open-sourced) Currently we have given support to OpenCode, KiloCode, OpenClaw, and Codex, most of you already have used at some point. No plugin. No config file. No IDE restart. You click ONEINFER, a local proxy intercepts your copilot's requests, translates the format, routes to your self hosted model, and returns ... ... expects an OpenAI messages format and your local model returns something else. Switch back to original models in one click, mid-session, no restart. For when you actually need it. This is just the start. Support for more agentic harnesses

... watchdog triggers a graceful shutdown when RSS crosses the threshold. It s useful as a safety net, but it was firing too eagerly during recoverable native-memory spikes. With the heap ceiling at 60% of the limit, V8 still hard-stops growth well below the watchdog, so the new threshold gives transient spikes more room to resolve before we preemptively cycle the pod. Confirming the fix against real kubelet The change was small, but justifying the rollout took more work ... ... risk cut the other way too: a too-tolerant liveness probe could mask a genuinely stuck pod and delay a restart that would actually help. We wrote a small reproduction harness

... relied more on raw authority signals. Now, context matters more. A link from a small blog that your exact customers read is worth more than a link from a big site that no one in your industry visits. The hard part is that DA is easy to measure. Relevance is not. You cannot put a number on "does this audience actually need what I sell?" So people ignore it. We track engagement from referral traffic, not just DA. If people ... ... conversation. From "can I get a link" to "can I help someone write a better article." The link follows. The trust follows. The old proxy is easier to measure. It is also easier to game. The new job is harder

... text. Even the general structure doesn't tell you much, at least for articles where humans were pushed into the same handful of SEO-friendly templates before generative AI became widespread. Comment from Fabrizio Pfannl(@fabriziowexare): Honest answer: AI-vs-human detection is mostly broken now. What still works is detecting whether a piece of writing was edited by someone with an opinion versus

... said "never again" to hardware.    I meant it. I had done it before. I know what hardware

... other person's perspective and figure out what to say next. In moments of self-doubt rather than offering endless validation, the app helps you identify patterns in how you relate to others, so you can actually grow. The hardest part of building this way is that success looks different. We can't optimize for "conversations with AI" because our whole thesis is that the most important conversations happen outside the app. We measure whether users report feeling more confident ... ... their real relationships which is harder