my current opinion is that people are over optimizing for multi-agent and under optimizing for compute allocation.
most agent failures i see aren't because there weren't enough agents, it's because the system treated every step equally.
reviewing shouldn't get the same budget as writing. planning shouldn't get the same budget as execution. retrieval shouldn't get the same budget as synthesis.
good agents probably don't need fixed thinking budgets or fixed topologies.
they need the ability to decide:
"is this worth thinking harder about?"
sometimes that means one model doing multiple passes.
sometimes that means spawning specialists.
sometimes that means stopping early.
static pipelines feel like compiling intelligence into infrastructure.
i think "compute allocation" is broader than just these you mentioned.
there's also allocation across:
context: what information deserves to stay live vs get compressed or dropped
verification: which outputs deserve checking vs trusting first pass
tool depth: when to search once vs branch into retrieval loops
parallelism: what should happen concurrently vs sequentially
memory: what should persist across steps vs remain local
latency budget: where users actually notice waiting
exploration vs commitment: when to generate alternatives vs lock in
a lot of systems quietly waste compute not on inference, but on doing expensive work too early or on the wrong representation.
multi agent is one way to spend compute.
adaptive control over where compute goes feels more fundamental.
to put it simply our agent should be smart enough to know when to ingest more and when to stop, when to loop, how much to take in, what questions to ask if not sure about something, what to do when a dead end occurs while keeping in mind the latency, tokens, hallucinations etc.
there are a lot of things that come into play when creating a docx file, and since you specifically said "format me a proposal", there are a lot of things taken into consideration of how each para should be styled/formatted.
if you see the docx skill claude has utilized you would know what it is doing and why, its utilizing js script to create a docx file for you.
how do you think claude will create a docx without using code?
i haven't been active on twitter these days so thought will give you guys a little update.
past months i haven't been able to learn something substantial but rather gave everything to my company (which i am proud of).
but now as things settle down i have made up my mind to post regularly and learn extensively.
i somehow really got interested in the ml field, so will be learning that for a while.
also i have revamped my portfolio: vedanshh.com
proof that i am alive and happy :)
Before the models, before the code there's math.
Revising linear algebra and calculus from scratch because I want to actually understand ML, not just use libraries.
The notebook says ML Engineer. I'm going to earn that title.
Red bottle. Purple keyboard. Let's go.
haven't touched core maths since class 10th (its been 6 years), this is gonna be tough - but i have never been this excited to learn something.
lessgoooo
465 Followers 3K FollowingBuilding autonomous software workflows. Deep in LangChain, CrewAI, and MERN stack. Sharing everything as I build in public. Let's ship. 🚀
246 Followers 487 FollowingSDE @OracleCloud | Specializing in networks and distributed systems | interested in Core CS and Psychology | CSE25 @IIITDelhi |
1.6M Followers 1K FollowingCo-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
16K Followers 358 Followingmain quest - @ravenmailinc. co-founder and cto. views == personal (my legal asked me to put this). https://t.co/b7RdLv6kjh ♥️
5K Followers 317 Followingbuilding @shellular_dev. run claude code, codex, cursor, terminals, git, and remotely access your computer's localhost without ngrok - all from your phone!
930 Followers 17 FollowingA boutique tech recruitment agency designed to bring you the best of Indian talent. Run by @_svs_, previously CTO at https://t.co/zYuZz66T90 and https://t.co/e6vHnLHK3t
1K Followers 442 FollowingOn a mission to prove you don't need privilege to succeed in tech | Building hardcore backend in Go | Microservices • gRPC • MLOPS • Kubernetes • AI
12K Followers 981 FollowingBuilding MVPs & Personal Brands
Founder @shadowrites
Designs @haarisuiux
Ghostwriting for founders & creators
Ex 2 US remote startups
DM for work & collabs