Elias Lumer

ai research & engineering

[Elias Lumer @EliasLumer]

@ishaansehgal If you have the agent running inside of the sandbox AND record every log/event it does, such that you can reply the session in a new sandbox, is there a reason why this isn’t also viable?

[Elias Lumer @EliasLumer]

@kunchenguid Have you thought about switching a model mid-task if u detect the agent not being sufficient enough for the task? Ex. A sidecar agent/jev watching the task to see if the agent is not getting enough context or not completing the task correctly?

[Elias Lumer @EliasLumer]

@devagrawal09 Isn’t the inclination towards Jev the same agentic workflow vs agent argument?

[Elias Lumer @EliasLumer]

@scheredev @steipete I’d imagine a lot of the complexity comes from data/api/iOS integrations from instinct

Re: innovation. The biggest innovation of the last year (coding agents) is just a simple LLM in a while loop calling tools, nothing complex about it

[Elias Lumer @EliasLumer]

Before, the industry turned to DAGs because they made agents more reliable/accurate (Langgraph)

Currently, the industry moved to agents because they can handle more and more dynamic complex inputs and uncertainty with high accuracy (Claude code , codex)

Future/now, the industry is looking back to DAGs as a cost & latency improvement rather than an accuracy improvement. If the task can be broken down into a DAG, it’s 100-300x cheaper/faster now.

Still I agree with you, short term memory, but good to put everything in perspective

[Elias Lumer @EliasLumer]

This works when the agent flow/user journey is simple.

Most times, need to handle:
1. Conversational queries + large workflows
2. Ambiguity of information, where the agent needs to spend a few minute figuring out the missing details and then creating a plan

Are you saying you can put that into code? And if so, what does it look like?

[Elias Lumer @EliasLumer]

@miu21590 @1yian I’m glad OpenAI started training their models without the reasoning effort/budget in the 3rd line of the system prompt for Astra . Update priors @devagrawal09 .

[Elias Lumer @EliasLumer]

@a33xb1 @pidotdev How do u like the quest layout? How does it remember things that span across quests?

[Elias Lumer @EliasLumer]

@Jacobbaggs3 @rakyll Did u look into it?

[Elias Lumer @EliasLumer]

@BillDA Where r u finding these files?

[Elias Lumer @EliasLumer]

@mstockton My thoughts too. Industry was forcing agents down everything, which worked because it was easy and prompt tuning got less important. Cost was always a factor but efficiency/latency is now a better indicator to switch some parts of the agent to deterministic JEV functions

[Elias Lumer @EliasLumer]

@HamelHusain General purpose classifier makes sense.

Not sure why we need new words either , other than to fuel the hype train haha

[Elias Lumer @EliasLumer]

@DhravyaShah R u releasing these .md’s? Do they have any structure to them that we can infer based on the md’s? Weekly vs daily structure would be interesting .

[Elias Lumer @EliasLumer]

@IamSaidSadik @DhravyaShah How so? What things can’t it do? And are other memory systems better/ for your use case?

[Elias Lumer @EliasLumer]

@DhravyaShah @mohsen____ Great stuff! I’m not following the git part. Is this local git? Or are user memories actually getting stored in a giant mono repo somewhere …

[Elias Lumer @EliasLumer]

@jerryjliu0 Is jev on your leaderboard? Curious on your benchmark accuracy vs latency for Jev vs others

[Elias Lumer @EliasLumer]

@CatalinCalistr1 @kushbhuwalka @typesafeai RAG usually means there’s some vector/cosine similarity involved . Technically “retrieval” could mean anything, but I was referring to vector cosine similarity. And when I said “rag is dead pt2” I was referring to it being killed by Grep/Bash and agentic coding/search

[Elias Lumer @EliasLumer]

@tamarajtran @typesafeai Any anecdotal evidence this is any good in practice? A cool use case but need to know if it holds up