Berkay M

We ran 21 raw LLMs as code reviewers on @withmartian's Code Review Bench: 50 real PRs, up to 3 effort levels, 3 passes, 7,980 reviews. The best raw models score as high as the best review tools. More thinking buys recall and costs precision. Every review, verdict and cost is open. Top score: Claude Sonnet 5.5 at low effort, F2 64.4, $0.14 per PR, 9 s per review. Best value: Muse Spark 1.3 at low effort reaches the top group for $0.06 per PR. Post: ironbee.ai/blog/the-honest-r… Code and data: github.com/ironbee-ai/code-r… We will announce @ironbee_ai's own code review tool very soon!

How we built a browser agent (at @ironbee_ai) where every step is one classifier call: @typesafeai's Jev picks, and an LLM steps in only when it matters. 9 actions in 6.7 s for $0.0005, plus everything that broke on the way: ironbee.ai/blog/how-we-built…

We just released @ironbee_ai Express, the FASTEST and the CHEAPEST browser agent powered by @typesafeai's Jev, with an LLM for the intelligence when it matters. A full checkout on our demo shop: 9 actions, 6.7 s, $0.00054 in model cost, review included. You don't write assertions. Jev checks the page, the API responses and the backend logs. If the page says "Order placed" but the API says FAILED, the run fails and shows why. Passing runs replay without step decisions: 20 actions in ~4 s, and it notices when the site changes. The code is on GitHub: github.com/ironbee-ai/ironbe… To run it on the IronBee platform, join the waitlist: ironbee.ai/express?utm_sourc…

A testing tool offers to run your app for you. All they need is your repo. Write down what they promised: env vars, feature flags, twelve services with real state, data at real shape, network latency, clock drift. A repository is the part of your system that fits in a box.

AI writes code up to 100x faster than it did two years ago. Checking whether that code actually works has not sped up at all. I felt it firsthand. I became the QA engineer for my own AI agent. It wrote the code, said "done" with rocket emojis, and I found the bugs it could not see. Screenshot, logs, repeat. That does not scale. So we built @ironbee_ai. The moment your coding agent finishes a change, IronBee runs it, watches how it behaves at runtime, and hands back a pass or fail with real evidence. Not a guess from reading the diff. Proof from running the code. Today @ironbee_ai is generally available. Public, free to start, open to everyone. Try it free today, and read why verification, not generation, is the real bottleneck of AI coding. Links in the comments. Start free: console.ironbee.ai How it works: ironbee.ai The full story: ironbee.medium.com/on-coding… #AI #DevTools #AIAgents #SoftwareEngineering

We benchmarked Browser DevTools MCP vs Playwright MCP on the same flow: → 78% fewer tokens → 76% fewer turns → 66% lower cost → 57% faster → Much more consistent run-to-run How we did it (tool design, execute batching, scoping): medium.com/@serkan_ozal/brow…

I read the ClickHouse blog post (clickhouse.com/blog/llm-obse…) where an AI assistant uses MCP to query OpenTelemetry data in ClickHouse and perform root cause analysis. It works to a point. Traces, metrics, and logs are great at localizing a problem. But in real production incidents, localization is not always the same thing as understanding the actual root cause. Once you know where the problem is, sometimes you need to go deeper. Application state. Runtime behavior. Host, VM, container, or Kubernetes level signals. That data usually isn’t collected continuously because it’s expensive and adds overhead. This is exactly the problem we are working on. Instead of collecting everything all the time, we use AI to decide when a problem happens and then collect deep, high fidelity data on demand across the full stack. Application, infrastructure, and environment. Safely, with guardrails and PII controls. The ClickHouse experiment shows what AI plus observability can already do. The next step is giving AI the ability to actively investigate systems when it matters, not just explain dashboards. That is where full stack autonomous debugging starts. #observability #opentelemetry #AI #debugging #SRE #production