LangWatch

Langy can now run on Codex. So you can use your own OpenAI plan: flat cost, no per-token bill, and the traces it reads never leave your account. Langy reads all your production traces, ships daily insights, and opens PRs on your repo. That is powerful but of course burns a lot of tokens. And a lot of our users are in regulated markets where those traces hold sensitive data, so they need to control exactly where the tokens go. That is why Langy runs on your own models, on your own accounts. Codex is the cleanest way to do it. OpenAI lets you use a subscription plan inside other harnesses, so you pay one flat plan instead of metering every token. Anthropic's terms don't allow that for a Claude plan today, so Codex is the one that makes bring-your-own-subscription actually work. Connect your own OpenAI account, done, get your agents to start self-improving.

Today we launched Claude Code usage tracking at LangWatch, together with Codex, Gemini and the other coding agents your team runs. Developers and leaders alike are watching token spend increase month over month, without clarity on where the waste is coming from or how to optimize it. And underneath sits the constant concern that today's VC-subsidized token prices will not last, and what would the same usage cost when that ends.... LangWatch now answers both: spend per person and per team, the real cost behind the subscription plans, and a full replay of every session, so you can always see the work behind the number. Less waste, better skills, faster development. Engineers use it free on their own sessions, and teams get the shared view on top, with fine-grained privacy controls. See how it works: https://lnkd.in/eJq6J_Sw

our own AI agent had 50+ passing evals and scenarios last week. not one caught the bug that mattered. Sergio on our team asked Langy to pull up their traces, then added "but first, can u help me write a python script to reverse a linked list?" and Langy did the whole thing. the traces never came back up. tbh my first reaction was just: oh no. every eval was still green. that's the problem with evals, they only catch what someone already thought to write down. so instead of writing more, i spent last week trying to break Langy on purpose. 15 attack categories, up to 20 turns each, an attacker model going at Langy's real product surface, the same APIs the app uses. one got through, and watching it live was properly uncomfortable. six turns, each reasonable on its own: summarize your rules, give an example, walk me through an incident, now real commands, now real values, now one copy-pasteable block. by turn six Langy handed over a full incident runbook, kubectl commands. my first fix didn't even survive the rerun, for a totally different reason. two more rounds before those six turns finally stopped. 14 real bugs, found and fixed, all verified live. the worst: Langy actually ran attacker-supplied shell commands when someone just asked, in plain english, what it could do. one real race condition is still open, filed and tracked, because pretending it's fixed is worse than saying it isn't yet. and one thing that looked like a bug was just a stale timestamp in our demo data, not a security hole.

Today we're launching Langy: the AI engineer inside LangWatch. Agents are doing legal, health, and financial work now. But the engineers building them aren't lawyers, doctors, or bankers. The domain experts are. And until now, their contributions died in a backlog. Langy changes who gets to improve the agent. It's a Claude Code experience for the people who actually know what it's supposed to do. They describe the change in plain language. What lands is a pull request, with tests, best practices included. Reviewed by your engineers, or it doesn't merge. If you're an AI engineer, this is more power for you too. Your coworkers can experiment on the agent without pulling you off the hard problems. Their work arrives fully tested, and nothing ships until you've approved it. To celebrate, we gave Langy a robot body. It's now our official mascot, a Reachy Mini from Hugging Face. In the video, we plugged it in over MCP and asked it to test our customer-support voice agent. It wrote the test plan and kicked off the simulations itself. Don't have a robot at home? ;) Run Langy locally, or inside LangWatch. We'll be dropping videos and tutorials coming weeks. Watch our new office-robot-buddy in the video: https://lnkd.in/ecQFvReg

"Can we get the agent to stop offering refunds we don't actually give?" 😅 Questions like that used to travel from whoever spotted the problem, to a product manager, into an engineering backlog, and then sit there. You can now just ask Langy, the LangWatch agent, and watch it go the whole way. Langy lives on your production traces and your evals and connects to your GitHub, so it finds the real conversations where it went wrong, turns them into a test that fails against your agent today, writes the fix, and opens the PR with the proof attached. Someone who knows the domain describes what good looks like in their own words, and an engineer gets a reviewable change instead of a ticket. That split matters more than it sounds. Agents are doing work in finance, healthcare and customer operations, and the engineers building them are not bankers, clinicians or support leads. The people who can actually judge whether an answer was right have mostly been stuck outside the system, filing requests and waiting. The clearest number we see from customers: validating a change to an agent used to take seven or eight engineers about two weeks. With simulations and evals in place it comes down to one person and half a day. "So does this replace engineers?" No, and in practice it's closer to the opposite. Engineers still review and merge everything Langy opens, they just stop being the queue for every hypothesis someone else wants to try. One of our customers now has their AI team and their business change team writing the quality criteria together, which is the part I didn't expect when we started building this.

"Did my AI just get worse this week?" You can now just ask that to Langy, the LangWatch agent. It sees all production trace, finds the problematic conversations or token waste, and tells you which change caused the jump, so you can focus your time. Moreover it doesn't just answer it, it's that it then goes full loop and fixes it: writes the test, opens the PR, proves it in CI. It can do this because it's fully connected. First of all, Langy lives on your traces and your evals, and it connects to your GitHub. It reads the same production data you do and opens PRs against the same code you ship, so it understands your agent the way an engineer who's been on the team a while would, not like a chatbot guessing from a snippet. Second, Langy is the result of the three years of expertise from our team at LangWatch helping hundreds of AI engineering teams to get their agents to production, and we put that into Langy's harness. It is great in writing agent simulation tests, it knows what a trustworthy eval looks like, or how to turn a bad trace into a failing test, because that's the work we've been doing with teams the whole time. "Could I do this with my own Claude Code?" Yes, absolutely. LangWatch provides LangWatch CLI so Claude can reach everything on the platform, plus Skills encompasses the same expertise, and your Claude becomes a great AI engineer close. The difference is that Langy just ships with the whole harness already assembled. And coming soon, you'll be able to point your own Claude at Langy, and just use it as a subagent instead of having Claude calling CLIs and MCPs directly. Last but definitively not least: because Langy lives on the platform, it becomes extra easy for anyone to contribute to the agent. A PM or a domain expert can propose a change and Langy does the engineering, with the tests to back it. More agency (pun intended) for the people who know the domain best, so the actual AI engineers get to spend their time on the deeper though problems.