Jimil Patel

Creator of AgentsPod 🎙️ I dig into the latest in agent engineering and turn it into a podcast, with AI as my co-producer. Episodes ↓

[Jimil Patel @jimilp7]

Agent engineering moves faster than anyone can read.

So I built an agent to keep up:
📄 new research papers
🛠️ engineering blogs from top labs
🧵 posts from the best builders

I turn the best into a 30-min podcast with two AI hosts.

This is AgentsPod: https://t.co/PXtZP0suuO

[Jimil Patel @jimilp7]

Sorry this isn’t answering your question - but a big part of model behavior is determined by model training - and sometimes we simply are at the mercy of what’s important to the labs. So the memory retreival patterns and such may not be meeting your expectations until models are trained well on this.

[Jimil Patel @jimilp7]

@JoshARosen Agreed - I also notice agents like brute forcing its way into finding information rather than starting narrow and widening the search. This happens to me for example if the agent is trying to get some info from a storage bucket where it will try to traverse every folder.

[Jimil Patel @jimilp7]

@Yuchenj_UW AI code, AI code everywhere

[Jimil Patel @jimilp7]

@TonyChanBuilds btw, does being verified help?

[Jimil Patel @jimilp7]

How are people managing context for different projects when using coding agents? I maintain a second brain that I can't work without at this point but see a lot of folks who don't have one.

[Jimil Patel @jimilp7]

@iam_3bu1 this is a good idea for cloud agents!

[Jimil Patel @jimilp7]

For people who run several claude code terminals at the same time working on different things - how do you maintain the code quality and engineering bar?

[Jimil Patel @jimilp7]

@pauliusztin_ Yeah - we should use multi agent is to offload and manage context wisely by letting other agents "explore" and return only relevant context back.

[Jimil Patel @jimilp7]

@harshpatel071 Prove you can make something valuable that is difficult to clone (By shipping something with real users)

[Jimil Patel @jimilp7]

@divaagurlxw Code is tokens, distribution is not.

[Jimil Patel @jimilp7]

@birk Building great cli for every resource/service. Enables lots of things like verification loops, context fetching, automating workflows etc. Turns out llms know how to use clis very well.

[Jimil Patel @jimilp7]

@yongfook Agreed - we are already seeing this in software engineering as well - there is complete AI SDLC where the human engineer’s job is to meet with people, get the requirements, get the business context etc as input to the agents who then deliver it e2e. Code seems solved.

[Jimil Patel @jimilp7]

@hwchase17 I don't believe so - as models get better - skills will be absorbed in the models training i.e llms will get better at common workflows - but memory will stay outside the models' weights so memory != skills.

[Jimil Patel @jimilp7]

@ShivamPatxl Thanks - but I’m talking about the step before that - the architecture design, code design etc. codex will only review what claude has written from a “finding bugs” perspective. Does it review whether the code can scale, factor in business logic etc?

[Jimil Patel @jimilp7]

Full episode (32 min), with chapters and every source:
https://t.co/4GEAOS6vOi

I dig into the stories and shape every episode, with AI as my co-producer. Follow along for more.

[Jimil Patel @jimilp7]

Same AI model. Can Bölük (@can1357), security researcher, changed only how his agent edits files. Success rate: 6.7% → 68.3%.

The model wasn't the problem. The tooling around it was.

Episode 1 of AgentsPod, my podcast on building with AI agents 🧵

[Jimil Patel @jimilp7]

Nicholas Carlini at Anthropic had 16 Claude agents build a C compiler. His biggest lesson: the tests that verify the agent's work have to be nearly perfect.

Easy way to check yours: break the code on purpose and see if your tests catch it.

https://t.co/Zs11tT9jdc