Rohan Mehta

@tamaybes Zuck owned 28% of a $100B company, and was ~the only founder with that much equity. Ant has 7 founders with 2% each of a $400B company. Don’t think your argument is correct.

It’s striking how little of the value of AI labs accrues to founders. Anthropic’s cofounders reportedly hold only ~1.8% each. Sam Altman had no equity in OpenAI for years. No founder of a frontier AI lab appears in the top 500 of the Forbes billionaires list, unless they made most of their wealth from other companies. This looks quite different from earlier tech companies. At Facebook’s IPO, Zuckerberg held ~28% of the company. At Google’s IPO, Larry and Sergey held ~16% each. I think this mostly reflects the fact that by far the most valuable factor of production in frontier AI is compute, not founders. Building a leading lab requires tens of billions in capital before meaningful revenue, and each raise dilutes founders substantially.

@matt_ambrogi @prashantmital @simonw If you use `result.to_input_list()`, it will take care of it for you. If you're passing in items manually, just make sure you include any reasoning items!

@prashantmital @simonw If I'm using the Agents SDK (which I know uses responses under the hood) but I'm managing my own conversation history, how can I make sure the reasoning is sent back in the message chain?

surprised and humbled by the reaction to my post on Responses API. @simonw’s QRT inspired me to continue unpacking how things work under the hood. so here’s a look at why Responses is better for agentic function calling in reasoning agents. 🧵👇 let me know what you think!

@TomasReimers @ankrgyl @jezell Oh I see what you mean. Technically yes you can call list_tools and then change the descriptions before passing that to an LLM, but not part of the spec.

@_rohanmehta @ankrgyl @jezell For the client to override them?

Spot on advice from @ankrgyl who is doing amazing things at Braintrust. You'll get a lot more mileage out of crafting custom tools for your LLM that understand how LLMs work and what the task at hand is than giving generic API tool slop to the LLM and trying to prompt engineer results. Coincidently, this is one of the reasons I'm not very optimistic about MCP. You don't want overly generic tools designed by people who have no idea what you are trying to get your agent to do. Might work ok for AI influencers to play around inChatGPT, but it's certainly not the best path for custom agents that need to accomplish specific tasks in production.

@TomasReimers @ankrgyl @jezell MCP already allows tool descriptions right? wdym?

@ankrgyl @jezell Oh wow, small world - I think that MCP will have to evolve to let you specify tool descriptions (at minimum). But in that case, he pointed out that APIs are probably a better abstraction anyway.

First, MCP. We’ve been really excited about MCP since it launched and it’s so good to see it land in the API. MCP allows you to connect your app with tons and tons of remote functionality from Stripe, Shopify, Cloudflare, and many more.

@stuffyokodraws You can configure the environment from chatgpt.com/codex/settings…, and add a setup script that has internet access!

Sadly it couldn't do a simple run lint command because it can't properly set up the env. I really wanted this to work since I use chatgpt so much

Trying out Codex How is it supposed to work if it doesn't have access to the internet? Like i can't install npm packages or upgrade things in my project

@rauchg @onequince @nextjs If 5 basis points=$59M, total revenue is $118B. That’s unlikely, guessing they actually meant 5%?

Improving web performance remains the closest thing that exists to a money printer. Amazing writeup from the @onequince engineering team on their @nextjs migration, which unlocked an incremental $59M annually and massive agility. I've now seen so many of these migrations. The coolest part, which they note in the article, is that Next.js is not just a 'faster horse', but provides the platform that enables future experimentation and personalization. Legacy static solutions, or even modern ones heavily reliant on full-page CDN caching, struggle with dynamic and AI-driven content. As someone in the weeds of the framework and the tech, I get really excited about these stories, because I know just how much *more* performance improvement is still up for grabs. tech.onequince.com/rebuilding-qui…

@omooretweets Next: "suggest 10 more songs I'd like" -> MCP server to create a spotify playlist

@haridigresses @AcquiredFM ooc, why is that?

@AcquiredFM This company needs to be outcompeted, sued, and antitrust-regulated out of existence. One of the most harmful companies in the country.

Tomorrow.

@hwchase17 IMO the intent here isn't "graphs are bad", it's needing to learn a DSL to declare your graph that isn't great (e.g. using custom conditionals/loops etc instead of just regular programming)

pretty bad advice here from openai there are 27 libraries like "Agents SDK" (i would put the original LangChain in this camp!) and none of them are reliable enough to get to production for 99% of use cases ✒️blog coming this weekend

@charliermarsh Do you use both Pycharm and Cursor?

For reference, with Cursor, I don't use the Chat (Composer?) or Agent modes at all. I just use tab completion and a lot of Command+K. It just lets me right the code I want much, much faster.

@ivanleomk Can you DM me the request ID? Should be in the error logs

Trying to get puppeteer to take snapshots with the agent-sdk and answer questions but getting weird issues that I've overflowed the context window for 4o. any suggestions @_rohanmehta ? An image shouldn't cause this :(

@Ed_Forson @OpenAI Ah good point and good feedback. We'll improve this.

@_rohanmehta @OpenAI I was just trying to write an application with an agent that does some web search + summarisation and then hand off to another agent that does TTS voice over. So in this case I don't need the STT input. Not sure how I could do this with the VoicePipeline paradigm.

Just gave a @OpenAI's Agent SDK a quick spin. Pros: - Good documentation - Very quick to get a working example - Handoffs, Context, and Callbacks are welcome constructs Cons: - Lots to learn still - Don't really like the VoiceAIPipeline stuff Early days.

@sama Docs/examples: openai.github.io/openai-agents-…

people love MCP and we are excited to add support across our products. available today in the agents SDK and support for chatgpt desktop app + responses api coming soon!

Docs here: openai.github.io/openai-agents-…

MCP support now available in the openai agents SDK! Especially love the tracing dashboard integration

great team, great product

@mikeysee @OpenAI @typescript Coming soon!

Bah @OpenAI why would you make your Agents SDK python only?? Show us @typescript users some love please :)

@swyx Sent out a PR to simplify this. Thanks for all the other feedback too - we're on it!

@_rohanmehta i wish the pretty printing for things were nicer out of the box. i have to manually indent this to read anything

about to code up a brand new project with full stack openai agents platform: - agents sdk - handoffs - guardrails - tool calling - tracing - web search - file search - (maaaybe) computer use so nice to use my own podcast to double check things lol. wish me luck. yes will open source. ive increasingly grown to dislike livestreaming because the focus should be on the code, not making Chat happy