आम्ही काय तयार करत आहोत आणि ते आतून कसे कार्य करते.
पोस्ट इंग्रजीमध्ये लिहिल्या आहेत. साइटचा उर्वरित भाग तुमच्या भाषेत उपलब्ध आहे.
Every agent we had built starts when a person types something. The seventh starts because it is Monday. That one change took away the approval dialog, the client-side tools, and the credential — and each of the three had to be replaced with something that works at 3am.
Google publishes no read-only scope for Crashlytics, so connecting asks for administrative access. We could not narrow the credential, so we narrowed everything else — and then found four bugs in ourselves getting there, including a header Google deletes in transit.
We are building one AI agent a day on our own platform. Five in, and each one found a bug nothing else could — a discarded system prompt, our own rate limiter, a missing PATH, and every media tool broken since it was written.
Most AI spreadsheet tools average a column that contains text and tell you the number with a straight face. Ours profiles the data first, says what is wrong with it, and writes charts as real files you can open.
Point it at a public repository and get an architecture note where every claim carries the file path it came from. It starts where execution starts, and it says which parts it did not read.
When a test fails, one of two things is wrong — and the code is the one nobody is protecting. Our Test Writer runs the tests it writes, and reports the bugs it finds instead of editing them away.
Transcribe, translate, speak — three model surfaces chained in a single agent run. Here is where the time goes, what it costs, and why the agent tells you what it cannot do before it spends anything.
The first agent we built that writes into somebody else's workspace. It turns messy notes into a short list of real tasks, and every write to Notion or Linear stops and waits for you to allow it.
The obvious way to let an AI agent edit audio is one tool that takes an ffmpeg command. It is a third of the code, it is remote code execution, and it makes the model worse. Here is what we built instead, and the two bugs we found on the way.
AI has been an English-language product for people who read English. We changed that: the whole app in three Indian languages, credits from ₹100, and a pipeline that adds the next language in minutes.
Buy AI credits with UPI, GPay, cards or netbanking — from ₹100 (about $1). Credits last a year, unlock every model, and there is no plan to cancel.
One cookies() call in a root layout opted our entire App Router tree out of static rendering. Here is how we found it, and four other Next.js 16 performance traps we hit fixing it.
आणखी येत आहे — मार्गदर्शक, मॉडेल तुलना आणि त्यांना चालवताना आम्ही काय शिकतो. RSS द्वारे सदस्यता घ्या.
