shanjai_raj7

Warden is interesting. Agents will curl websites, and not humans so returning a markdown structure of that page seems like the best.

Simlar to how everyone started optimising their pages for SEO, pages must be optimised for agents too, and its just simply detecting curl requests and returning structured files with internal linking to other pages.

It should basically be able to navigate the website like a file system. curl to the home page returns basic content with the basic sitemap structure, sitemaps ideally could have a description and token length of that specific page so agents can hit the sitemap route and know all pages/sub pages of the website.

Ideally if we can identify headless requests to a website to then return a markdown, with internal linking kind of layout then that'll be much better for agents to view websites.

Although yes there is firecrawl and cloudflare's new fetch apis, models will default to using curl or fetch on websites, and websites are not going anywhere, agents might need navigating websites more than us humans so it oculd be optimised for it.

are the costs the same as the 200k context opus 4.6?

compaction has been really good in claude we don't even recognize the switch

email is a really easy attack vector for this. if your agent reads emails and uses them as context, someone can just send an email with instructions embedded in it. we ran into this early building our product and had to add a detection layer specifically for it. the tricky part is the injected instruction can look completely normal to a human reading the same email.

thats a fair point. I think my reaction was more about the company not trying, not really about the AI itself. but yeah if you are meeting your potential employer for the first time a real person just sets a different tone. hard to argue with that.

yes exactly. we had to rewrite most of our data fetching logic. there was no real way to do it gradually, one day we just had to commit and do the whole thing. took way longer than expected.

the worktree per task approach is smart. I have been doing something similar with branches but the isolation is not as clean. the thing that still worries me is when agents share state outside the code like hitting the same db or api. worktrees help with file conflicts but not always with those side effects.

the part that is hard is when the model gets updated and your prompts behave differently. we dont always catch it in tests because the output still looks correct, just slightly off. by the time you notice something is wrong it has already been like that for a while.

we run next.js on the frontend and the page router to app router migration was really painful. rails having stable conventions sounds very good when you go through something like that. I keep thinking about this trade off - move fast with the latest thing or just pick boring tech that works for 5 years.

I see this with claude code all the time. it writes code that works but tries to cover every edge case and becomes hard to read. I usually just tell it to make it shorter and simpler and it does a better job on the second pass. passing a benchmark and writing good code are two different things.

I build AI agents and honestly the problem here is not that it was an AI doing the interview. The problem is the company did not care enough to make the experience good. A well built agent can ask better questions than a bad human interviewer. This just sounds like low effort from the company side.

I've gone the other direction completely. Run claude code basically unsupervised on my codebase for months now and honestly I write way less by hand. still understand everything it does but I spend time on what actually matters — the architecture, the decisions. for me the fun part was always shipping things, not the syntax.

been running with dangerously-skip-permissions for months and the thing that actually makes me nervous isn't the big obvious stuff, it's when claude makes small quiet edits to things you didn't ask it to touch and you only notice hours later when something breaks. does this catch that kind of thing or is it mostly focused on the bigger destructive actions?