I wrote a book for engineering leaders. The Agentic Engineering Guide. 10 parts, 33 chapters. Context engineering, authorization, cost control, sandboxing, evaluation, governance, adoption playbooks. Not a prompt tutorial. The decisions you face when shipping agents to production this quarter. Early version. Open source. Rough edges included. Spot a mistake, open a PR. agents.siddhantkhare.com

irrespective of the agents and all, review in general is one of the biggest bottleneck when you have to move fast

Some thoughts & lessons on coding agents' cache losses - siddhantkhare.com/writing/ag…

Raw & unscripted video that i recorded late last night while i was awake to catch my early morning flight. tl;dr; whoami siddhantkhare I build the boring parts that make AI agents work piped.video/_QEbLePjHgg?si=DMPW…

I spotted GitHub Copilot Desktop using SQL to track its own tasks. In a separate SDK test, I could read the updated tasks after compaction & a clean runtime restart, before sending another prompt siddhantkhare.com/writing/wh…

I’ll be in New Delhi for the next two days. DM if you wanna catch up & chat!!

Genuine question: How much are these frontier labs paying these middle-human agents to do this kind of work?

My friend @bhawna_Pawar19 is building KarakCrew ☕ For those “I’d go, but who with?” moments. Find people nearby for karak, padel, hikes or a sunset walk. Join a plan or start your own. Friends, not dates. Free to use!!! karakcrew.com

They count themselves in the top 5%, but still struggle to have basic common sense to move “Zone/Seq” wise while boarding, and the same nonsense of getting up from their seats even when the Air bridge is not even connected

github1s now supports AI features. You can configure an LLM and chat with references to your code to help you understand it quickly. All data stays local in your browser. github.com/conwnet/github1s

It's very great to see some orgs are burning so much money to do sort of anti-Anthropic marketing on social media , but still, their product & engineering team still keep shipping and improving models.

A SHA-256 hash proves your trace wasn't tampered with. It says nothing about whether it captured everything that happened. Read about the blind spots in agent observability — a hook only sees what its host sends it, cost estimates can be off by 30-50%, and without labeling what's observed vs. inferred vs. missing, a "complete" trace is just false confidence with a checksum. Read the article by @Siddhant_K_code below: siddhantkhare.com/writing/a-…