Jamu Kakar
jkakar.bsky.social
Jamu Kakar
@jkakar.bsky.social
310 followers 540 following 160 posts
Cofounder of Blaide. Audio chisel. Slow clapper. Foot stomper.
Posts Media Videos Starter Packs
It’s another case with agents where it’s kind of neat, but not immediately practical. My use case is a personalized onboarding flow for an app, and I don’t think anyone wants to wait 30-40s between questions and do that five or seven times before they can start trying the app.
I had to force the agent to use the tools I built to evaluate answers and generate questions, because otherwise I couldn’t get it to use them purely with my prompts. It worked a couple times, but it was flaky.
I built an interview agent yesterday. You give it an initial question, a list of things you want to know, and a maximum number of question/answer rounds. It works but is painfully slow with gpt-5 with 30-40s waits between questions. gpt-4.1-mini is faster with 5-8s waits but the quality is poor.
I can imagine a travel site interviewing a user about a trip they want to book, for example, and then an agent built and operated by that site uses internal APIs to prepare options they can review and book. That seems much more plausible than a generic travel planning agent working in the wild.
I assume companies are going to want to maintain control, as we see with APIs today, to minimize fraud if nothing else. It’s hard to imagine a future of agents working autonomously, so I wonder if the tech will end up being used behind the scenes more than something exposes to users.
I imagine a future where this works will need online stores and other sites (travel sites, for example) to expose MCP-like facilities for agents. I assume they’ll need agents to be registered. Getting API access to a lot of sites today is either nonexistent or a painful and slow process.
When I did have success it was because I chose a smaller site without bot protections, and then oriented the agent workflow around that site’s specific behaviours. Making a generic agent that can work with any site seems like it’d be quite an undertaking.
Bigger sites I tried had bot protections that kicked in pretty quickly and broke the agent. Many sites require login before anything can be added to a cart. Pop ups and other surprises can confuse and break the agent’s progress. Even if it succeeds there’s no good way to get a user to a cart.
I spent a couple days cooking up an AI agent to drive a browser to try to add ingredients to an online grocery store shopping cart. I got it to work, but it’s fragile and takes 15-20 minutes to add 6 items to a cart, and costs about $1.25 in tokens. The hype is much more impressive than the reality.
You’re absolutely right!
Yeah, that’s a good point. It’s seen as a way to offload thinking and understanding vs being a tool to incorporate into those activities.
It’s fascinating to me that people hear about coding with AI and then think that it eliminates the need for the things we’ve collectively learned are needed to make software work (design, testing, refactoring, data modelling, contracts, etc.). The widespread magical thinking is wild to see.
I've been building Cookwith which is an app to generate recipes and meal plans tailored to users dietary needs and preferences (eg, protein goals, allergy considerations, etc.). I used this tofu marinade recipe for a party I threw yesterday and everyone loved it! cookwith.co/recipe/4fdab...
Cookwith - Greek-Inspired Tofu Marinade with Fresh Rosemary
This Greek-Inspired Tofu Marinade is a celebration of Mediterranean flavors that transforms tofu into a savory, zesty main. With the switch to fresh rosema...
cookwith.co
A couple months ago we went to the Studio Ghibli museum in Tokyo and it was sooooo cool. If you get an opportunity to be in that part of the world I recommend checking it out.
Yeah, same. Multitasking is neither healthy nor productive.
How do we square the fact that, “the mind can only do one thing at a time”, with a world of AI tools and expectations that an individual’s (or a team’s) output is going to be dramatically increased. I struggle to see how this won’t exacerbate burnout for all but the best meditators.
Strong agree! I’d love to see all cars banned with the exception of people with mobility issues, deliveries, and maintenance workers. I bet you could do that and change half the roads to pedestrian only spaces, which would improve the experience.
Waiting for the AI is the new "My code's compiling."
Compiling
xkcd.com
If I ever open an AirBnB it’s going to have the best toothpaste.
Yeah, I’ve observed that more than once, and I’ve always been grateful for it and find it inspiring.
Yeah. I was told that there was one other higher priority case than this one, but it still took awhile. The 911 attendant was apologetic about it. The paramedics were also great. I’m always impressed by people with these kinds of jobs. The level of professionalism is always remarkable.
Yeah, they were conscious. There was a language barrier, and they couldn’t say much for awhile, but their breathing was okay the whole time, and they didn’t break out into a sweat or anything. I’m going to check on them tomorrow and see if they’re okay.