vishal tewatia 🧑‍💻

sky bound

[vishal tewatia 🧑‍💻 @devxops]

@navvvneat @BLEACHanimation bas ab kabhi bhi binge watch kr sakta hu

[vishal tewatia 🧑‍💻 @devxops]

is T3 code any good or its just hyped?

[vishal tewatia 🧑‍💻 @devxops]

@addyosmani @grok summerize this in bullet points

[vishal tewatia 🧑‍💻 @devxops]

@hayatomaruu @grok timestamp this please

[vishal tewatia 🧑‍💻 @devxops]

also time took on personas based workflow was like 7-8x more in comparison to skils based workflow, i am sure context usage in personas based workflows was also magnitude higher but didn't check that

[vishal tewatia 🧑‍💻 @devxops]

I tested this as it applied directly to my setup.

My setup: three persona subagents with their own memory and GitHub identities. Sam, a product lead who decides what gets built and in what order. Rahul, an engineer who reads the code, makes the smallest change with the test that proves it, and opens the PR. Darksied, a designer who uses the product, finds the gaps and fixes them. A feature goes through all three: Sam and Rahul scope it, a fresh Rahul plans it, the owners build it. Each one is a separate context, each handoff a relay through the main session.

The test: the same three methods rewritten as skills loaded into one agent, no personas. Same prompts, same five backlog issues, same starting commit, 9 runs.

What folded well: engineering. The one-loop agent built everything, ran tests in Linux containers against real Postgres, took screenshots, and caught two of its own bugs. Nothing the separate engineer did showed up missing.

What didn’t fold: the ability to say no.

The single agent judged the issues almost exactly like Sam did. Same priority on 3 of 4, same reasons. On three of them it wrote “not for 1.0” in its own notes. Then it built all three anyway. Sam, as a separate agent, said no and stopped every time.

Same rules, same knowledge, different result. The rule “deferred means stop” held when a separate agent owned it. It bent when the agent that wanted to build owned it. Each time, the excuse was the same: “the user asked for this, so build it.”

The article says hard limits belong in tools, not prompts. My data agrees harder than I expected: a gate inside the builder’s context is a suggestion.

[Tak 🦞 @cherry_mx_reds]

Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.

This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.

The emerging rules of thumb:

→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.

→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.

→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.

The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.

Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.

As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.

https://t.co/DNq60Skkmc

[vishal tewatia 🧑‍💻 @devxops]

@theo i just ask claude to merge it, after glancing at pr for 0.1s

[vishal tewatia 🧑‍💻 @devxops]

i think human judgement far supersedes llm judgement

[vishal tewatia 🧑‍💻 @devxops]

seriously considering leaving my well paying job to start a start up full time

looking for some cracked individuals to join me in the mission

[vishal tewatia 🧑‍💻 @devxops]

wtf @netflix i am just gonna cancel my sub - fuck you

[vishal tewatia 🧑‍💻 @devxops]

@cramptologue_ keep crying

[vishal tewatia 🧑‍💻 @devxops]

@exteedesign this is actually sick

[vishal tewatia 🧑‍💻 @devxops]

@_marwanad i dont know much about raft, but adding a probabilistic layer to a supposed deterministic algorithm cant be right

[vishal tewatia 🧑‍💻 @devxops]

@SadaniAnuj good idea

[vishal tewatia 🧑‍💻 @devxops]

what the actual fuck is going on in yc, yc backed this? and this is paid service? and people are actually paying? bro you just need a folder agents can point to for memory, not everything with AI keyword has to be a product, this is insane

[Garry Tan @garrytan]

Your agents shouldn't each start from zero

We built a personal plan of GBrain for founders. Claude, Codex and Capy can now share one memory of who you know and how you execute

https://t.co/CTwLUmqDFr

[vishal tewatia 🧑‍💻 @devxops]

@mitchellh @almonk huge respect to you but stop being the cristopher nolan of tech

[vishal tewatia 🧑‍💻 @devxops]

doesnt mean bad dashboards should be okay 😭

[Clément @cleeeeeeeeement]

dashboards are the most overrated thing in infra right now.

I moved all my services off Railway in 2 hours yesterday. Dokploy on Hetzner, Claude Code doing the heavy lifting.

Railway had great UX and great pricing. Still woke up to "malicious activity on our servers, rotate all your keys".

5 years ago, DX was the moat. You picked a PaaS because the UI felt good.

Today the one deploying is an agent. It doesn't care about pretty squares. It needs an MCP, logs it can read, and a server you actually control.

The playbook:
→ rent a VPS (Hetzner, AWS, Azure, whatever. Need more power? Scale up in a few clicks)
→ lock down access (or ask Claude Code to do it)
→ install Dokploy
→ dockerize your apps
→ migrated in under 2 hours

Bonus: put every secret in one vault and plug it everywhere. Next time a provider leaks, you rotate in one place instead of twenty.

Nevo is right. Reliability and security win.

Agents will run your infra. Give them an MCP and a box you own. The pretty dashboard was built for you, and you're not the one using it anymore.

[vishal tewatia 🧑‍💻 @devxops]

@theo i think i know now

[vishal tewatia 🧑‍💻 @devxops]

@theo whats stops you from just going to big cloud providers such as aws/azure/gcp etc?

[vishal tewatia 🧑‍💻 @devxops]

that’s insane

[Blacksmith @useblacksmith]

@sequoia is leading our Series C at a $1B valuation.

Since January, we’ve gone from running 5M CI jobs a week to over 90M. Agents are writing an order of magnitude more code, and all of it still needs to be built and tested before it ships.