Sonup Bhuyaan (@imsonup_) header
Sonup Bhuyaan (@imsonup_) profile picture

Building AI agents in public. Sharing what breaks, what works, and the numbers behind both. Founder @Blevoapp

Joined April 2012
Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)Photo by Sonup Bhuyaan (@imsonup_)
The future of fitness isn’t about hitting the same numbers as everyone else. It’s about understanding your body, your habits, and your progress. That’s the future we’re building with Blevo. Get early access: blevo.app
10,000 steps a day started as a marketing slogan. In the 1960s a Japanese pedometer was sold as "manpo-kei". Literally, "10,000-step meter". Your body never signed off on that number. Join the Blevo waitlist: blevo.app
1
1
61
My health is everywhere. Steps in one app, test reports lost in my gallery, a diet plan I haven't opened in months. So we built Blevo, with an AI coach that reads all of it. Ask it "am I actually getting healthier?" and get a real answer. Log lunch by typing "do roti aur ek katori dal". It tells you each morning whether to push, go easy or recover. Built in India. iPhone and Android in December. First 500 on the waitlist get early access: blevo.app
1
1
35
Sonup Bhuyaan retweeted
Join waiting list blevo.app
1
2
103
Sonup Bhuyaan retweeted
What if your next game studio was just one prompt away? A few hours with Fable 5.2—and this is already looking like an AAA game.
1
2
140
a stuck agent looks exactly like a working agent from the outside. the fix is not a smarter model, it is a dashboard that shows progress per step
1
4
108
writing rules once and letting every agent inherit them beats another meeting summary. do you version them or just edit in place
Claude Meetings starts on Team and Enterprise, and I work solo. The idea behind it still lands for me. Most of getting good work out of an agent is closing the context gap. Mine is the rules I'd otherwise repeat, so I write them down once and my agents keep them.
1
83
the 2019 confluence page is the real prompt. nobody owns deleting stale docs, so who does
Throwing an LLM and a vector database at terrible, unstructured company docs doesn't make your organization smarter. It just gives you confident, hallucinated answers sourced from outdated 2019 Confluence pages.
1
77
hot take: the best agent builders in 2026 are not the best coders. they are the best at writing down what done looks like. agree or disagree?
1
56
the usb-c analogy works until an agent tries to use three tools in one action and the failure modes multiply. builders who ship tool use first need to ship the error budget and the rollback. which builder is measuring agent tool-use success rate in production right now
Everyone is chasing smarter AI models. But the real breakthrough is letting AI use your tools. MCP is like USB-C for AI. One protocol to connect agents with Slack, GitHub, Google Drive, databases and more. The future isn't AI that just answers questions. It's AI that gets work done.
1
48
right until the moment an agent has to decide whether to call another tool. the incentive flips the second you charge per action instead of per thought. who is actually building the cost-of-mistake accounting for agentic loops?"
I think AI labs don't have the incentive to deliver the best products for their users. Their pricing model is dollars per token, or a usage limit. That's why we get loops, agentic teams, dead skills and MCPs sitting in context, CLAUDE.md files stacking up.
33
every agent needs a budget before it needs memory. the runs that can't say stop are the ones that burn the money. what is your hard limit per run?
1
81
What if your next game studio was just one prompt away? A few hours with Fable 5.2—and this is already looking like an AAA game.
1
2
140
simulating the user before shipping is underrated. most agents are only tested on the happy path they were written for
Agent evals need to feel more real-world. Pretty cool to see @pluraiAI ship Plurai Simulation today. You get a digital twin of your agent's world and can test it there instead of in front of users. Live now and free to get started.
42
browser agents stop being demos the day they can say i am stuck instead of guessing
42
benchmark bragging rights aside, the thing that matters for agents is whether the sim-to-real gap stays small when the task actually moves. video pretraining for control is the right bet if the demo data isn't lying to the model about friction and slip
Replying to @runwayml
Praxis-1 dramatically outperforms frontier LLMs on speed and cost. On the same station and tasks, Praxis-1 finished 15.6x faster than GPT-6 Astra driving the arm, at 51x lower cost per completed task.
42
the real split isn't local versus cloud, it's who pays the token bill. a mac mini orchestrating an api is still an agent, it just routes the work instead of doing the inference. for most builders that's the more honest architecture
Replying to @jrs25_SD @pravijn
Mac mini are not fit for running most capable models locally - it’s only useful for orchestrating api based models (openrouter, or direct to a closed model company) through a harness (Hermes, Deepseek, etc) You really need a m5u studio or a dgx spark (or similar gb10) with more unified ram, and even then you are only scratching the surface I have 4 sparks (512gb total ram) where I run glm 5.3 flash, Deepseek 4.1 flash, and qwen models and I prob would prob just be running openrouter if I hadn’t gotten them at at original msrp. And I do a lot of experimentation (fine tuning, etc). If I was just running inference I prob would just use openrouter still today.
1
40
indian builders shipping ai agents in public right now. follow all of them @itsumeshk: runable grow, an agent that runs your whole go to market @ShubhamInTech: agnost, finds where agents silently fail @an_engineer_log: agent37, agent sandboxes for $4.76 a month @SaiVamshi1818: godhands, agents that learn by watching your best people @prukalpa: open sourced ai registry, one home for agent skills @chetan_sandeep: ai agents on whatsapp in under a minute @anushkmittal: shapes, multiplayer agents for group chats @Aiswarya_Sankar: agentplay, replay what your coding agent actually did @NandySomdipto: dani, the agent running dan lab's ops who did i miss?
1
6
160
the agent automated the posting. nobody automated picking the niche. that is still the whole job
Nas últimas 2 semanas, criei 3 contas em nichos diferentes no instagram, pra vender meus aplicativos de forma 100% orgânica. Desde então: > +15 milhões de visualizações > 70k+ seguidores novos > centenas de trials gerados de forma orgânica Insta 1: 24 videos por dia no pico (1 por hora) Insta 2: 14 videos por dia no pico Insta 3: 11 videos por dia no pico Tudo isso 100% automatizado usando Claude Code + Eleven labs enquanto trabalhava em outras coisas. Ah e isso aqui foi completamento orgânico, 100%. Eu estava há meses buscando estratégias que funcionassem no orgânico e acho que finalmente encontrei.
1
47