🗣️ Post 1: Show HN: Open-source alternative to ChatGPT Agents for browsing

As posted by: ElasticBottle  |  🔥 Points: 81

https://github.com/trymeka/agent

💬 Summary

Hey HN,

We are Winston, Edward, and James, and we built Meka Agent, an open-source framework that lets vision-based LLMs execute tasks directly on a computer, just like a person would.

Backstory:

In the last few months, we've been building computer-use agents that have been used by various teams for QA testing, but realized that the underlying browsing frameworks aren't quite good enough yet.

As such, we've been working on a browsing agent.

We achieved 72.7% on WebArena compared to the previous state of the art set by OpenAI's new ChatGPT agent at 65.4%. You can read more about it here: https://github.com/trymeka/webarena_evals.

Today, we are open sourcing Meka, our state of the art agent, to allow anyone to build their own powerful, vision-based agents from scratch. We provide the groundwork for the hard parts, so you don't have to:

• True vision-based control: Meka doesn't just read HTML. It looks at the screen, identifies interactive elements, and decides where to click, type, and scroll.
• Full computer access: It's not sandboxed in a browser. Meka operates with OS-level controls, allowing it to handle system dialogues, file uploads, and other interactions that browser-only automation tools can't.
• Extensible by design: We've made it easy to plug in your own LLMs and computer providers.
• State-of-the-art performance: 72.7% on WebArena

Our goal is to enable developers to create repeatable, robust tasks on any computer just by prompting an agent, without worrying about the implementation details.

We’d love to get your feedback on how this tool could fit into your automation workflows. Try it out and let us know what you think.

You can find the repo on GitHub and get started quickly with our hosted platform, https://app.withmeka.com/.

Thanks, Winston, Edward, and James

🗣️ Post 2: Show HN: An AI agent that learns your product and guides your users

As posted by: pancomplex  |  🔥 Points: 60

https://frigade.ai

💬 Summary

Hey HN! My name is Christian, and I’m the co-founder of https://frigade.ai. We’ve built an AI agent that automatically learns how to use any web-based product, and in turn guides users directly in the UI, automatically generates documentation, and even takes actions on a user’s behalf. Think of it as Clippy from the old MS Office. But on steroids. And actually helpful.

You can see the agent and tool-calling SDK in action here: https://www.youtube.com/watch?v=UPe0t3A1Vpg

How is this different from other AI customer support products?

Most AI "copilots" are really just glorified chatbots. They skim your help center and spit out some nonspecific bullet points. Basically some ‘hopes and prayers’ that your users will figure it out. Ultimately, this puts the burden on the user to follow through. And assumes companies are keeping their help center up-to-date with every product change. That means constant screenshots of new product UI or features for accurate instructions.These solutions leverage only a fraction of what’s possible with AI, which can now reason about software interfaces extensively.

With Frigade AI, we guide the user directly in the product and build on-demand tours based on the current user’s state and context. The agents can also take actions immediately on a user’s behalf, e.g. inviting a colleague to a workspace or retrieving billing information (via our tool calling SDK).

This was only made possible recently. The latest frontier models (GPT 4.1, Claude 4, Gemini 2.5, etc.) are able to reason about UIs and workflows in a way that simply didn’t work just 6 months ago. That’s why we’re so excited to bring this technology to the forefront of complex legacy SaaS applications that are not yet AI enabled.

How does it work?

1. Invite agent@frigade.ai to your product. You can send multiple invitations based on distinct roles.
2. Our agent automatically explores and reasons about your application.
3. Attach any existing help center resources or training documentation to supplement the agent’s understanding. Totally optional.
4. Install the agent assistant Javascript snippet (just a few lines).
5. That’s it. Your users can now start asking questions and get on demand product tours and questions answered in real time without any overhead.

This process takes only a few minutes. Once running, you can improve the agent by rating and providing feedback to the responses it provides. If you want to integrate further, you can also hook up your own code to our tool calling SDK to enable the agent to look up customer info, issue refunds, etc. directly. These calls can be made with just a few lines of code by describing the tool and its parameters in natural language and passing a single Javascript promise (e.g. make an API call, call a function in your app, etc.).

Would love to hear what the HN crowd thinks about this approach! Are you building your own AI agent from scratch, or looking to embed one off the shelf?

🗣️ Post 3: Show HN: OpenAI Agents SDK demos made durable and scalable with Temporal

As posted by: stevea-temporal  |  🔥 Points: 12

https://github.com/steveandroulakis/openai-agents-demos

💬 Summary

Steve from Temporal here. Temporal is an MIT open source project for reliable execution at scale. I adapted+extended some of OpenAI's Agents SDK samples to integrate with Temporal.

These demo agents can survive process crashes, scale to millions of executions in parallel and have easy-to-implement human interactivity. Just add a couple of Python decorators to your OpenAI agent code, run Temporal workers and you're ready to go.

Check the video I did with OpenAI showing this in action (it's linked in the repo).

OpenAI actually use us for ChatGPT Images and also their Codex code writing agent so I figure this may be a viable path for others to code something easy that's also reliable at scale.

Happy to answer questions.

🗣️ Post 4: Reddit and Perplexity got us leads faster than Google ever did

As posted by: graveEra  |  🔥 Points: 8

https://news.ycombinator.com/item?id=44737677

💬 Summary

We used to play the SEO game.

Write blog

Wait 3 months

Maybe rank

Maybe convert

What actually happened at 8 early-stage AI startups (Series A or earlier):

-Google Page 1 took ~94 days

- Organic CTR: 2.6%

- First qualified lead: 6–8 weeks

So we ditched the playbook.

We asked one question: How fast can we show up when someone asks ChatGPT or Perplexity what tool to use?

Turns out… faster than Google. And yeah, it brought pipeline.

What changed when we went LLM-first:

- Perplexity picked up our content in under 48 hours

- ChatGPT (with Browsing) indexed feature pages in 3 days

- 18.2% of sessions now come from LLM-originated paths

- Those leads convert 2.4x better than blog traffic

Then Reddit unlocked another level.

We posted no-link, technical breakdowns here.

One of them (about how we automated an AI agent pipeline) got quoted by Perplexity in:

- “UX AI Agent”

- “Best Firecrawl alternatives”

- “How to track LLM bots”

No push. No SEO. Just built in public.

3 days later:

- 9 Perplexity query quotes

- 2 inbound leads mentioned us directly

Reddit is training data goldmine for LLMs.

Here's what worked for us:

1. Add Q&A blocks to product pages (all <40 words)

   Example: Q: How does FireGEO detect ClaudeBot?

   A: It fingerprints known Anthropic headers and reverse-DNS matches IP blocks like 2600:1f18::/32

   - Indexed by Perplexity in <48 hours

   - 11 bot hits in 5 days

   - 1 lead → trial signup in <1 week

2. Build an ai-sitemap.xml

   Only high-signal pages: - API docs

   - Feature comparisons

   - Pricing breakdowns

   - Tech specs

   Crawl rate = 2.3x higher than default sitemap.

   GPTBot, ClaudeBot, and PerplexityBot show up daily in logs.

3. Treat Reddit as an input layer

   We post raw content here before it hits our blog.

   In last 30 days:

   - ~30,000 views across Reddit

   - 9 quotes in Perplexity answers

   - 2 leads directly from those mentions

If you’re shipping something real, try this:

- Install FireGEO or track LLM bots via reverse DNS + ASN logs

- Create llm.txt for structured answers

- Tag LLM traffic with UTMs and route to CRM

Curious to know what’s working for you around LLM visibility?

Any tactics or insights others here are seeing?

🗣️ Post 5: Has AI coding gone too far? I feel like I'm losing control of my own projects

As posted by: Shaun0  |  🔥 Points: 4

https://news.ycombinator.com/item?id=44743115

💬 Summary

I wanted to share some thoughts on AI coding assistants that have been bothering me for a while, and I think the analogy of "a kid with a credit card" perfectly captures the danger of what some call "vibecoding." At least until we have true AGI, this feels like a serious issue.

After using Cursor intensively for the better part of a year, I'm stunned by how fast it is. It can scaffold entire features, wire up components, and write complex logic in seconds. The feeling is like the difference between driving a car with a manual versus an automatic transmission. Or maybe, more accurately, like the difference between reading detailed documentation versus just watching a summary video.

It's brought me back to when I first started using GitHub Copilot in 2023. Back then, it was mostly for autocompleting methods and providing in-context suggestions. That level of assistance felt just right. For more complex problems, I'd consciously switch contexts and ask a web-based AI like ChatGPT. I was still the one driving.

But tools like Cursor have changed the dynamic entirely. They are so proactive that they're stripping me of the habit of thinking deeply about the business logic. It's not that I've lost the ability to think, but I'm losing the ingrained, subconscious behavior of doing it. I'm no longer forced to hold the entire architecture in my head.

This is leading to a progressively weaker sense of ownership over the project. The workflow becomes:

Tell the AI to write a function.

Debug and test it.

Tell the AI to write the next function that connects to it.

Rinse and repeat. While fast, I end up with a series of black boxes I've prompted into existence. My role shifts from "I know what I'm building" to "I know what I want." There's a subtle but crucial difference. I'm becoming a project manager directing an AI intern, not an engineer crafting a solution.

This is detrimental for both the individual developer and the long-term health of a project. If everyone on the team adopts this workflow, who truly understands the full picture?

Here’s a concrete example that illustrates my point perfectly: writing git commit messages.

Every time I commit, I have a personal rule to review all changed files and write the commit message myself, in my own words. This forces me to synthesize the changes and solidifies my understanding of the project's state at that specific point in time. It keeps my sense of control strong.

If I were to let an AI auto-generate the commit message from the diff, I might save a few minutes. But a month later, looking back, I’d have no real memory or context for that commit. It would just be a technically accurate but soulless log entry.

I worry that by optimizing for short-term speed, we're sacrificing long-term understanding and control.

Is anyone else feeling this tension? How are you balancing the incredible power of these tools with the need to remain the master of your own codebase?

🎯 Final Takeaways

These discussions reveal how developers think about emerging AI trends, tool usage, and practical innovation. Take inspiration from these community insights to level up your own development or prompt workflows.