Sid Uppal
@upster
early adopter; my thoughts, ever evolving.
You've heard @satyanadella say 30% of Microsoft code is written by AI. But agents are also speeding up dev workflows like PR completion, now 47% faster at P90. Big wins for productivity and dev-happiness. Learn more & how to build your own Teams agent 👇🔗
With MCP/A2A, the LLM issues a tool-call your code handles and responds to. With CUA, you're relying on the LLM's safety checks to pause before clicking or typing something risky, but those can be tricked. Fun challenges to solve.
Today we launched a new product called ChatGPT Agent. Agent represents a new level of capability for AI systems and can accomplish some remarkable, complex tasks for you using its own computer. It combines the spirit of Deep Research and Operator, but is more powerful than that…
You are a manager, and you are a manager, and you are a manager! 🤖
Learning AI has become table stakes for your career. The next competitive edge will be knowing how to manage a team of AIs.
The "30% of code by AI" stat misses a big win: internal loop acceleration. Folks are speeding up on-call investigations with GH Copilot + custom MCPs, automating repetitive toil, etc.
In podcast form, thanks to @NotebookLM: notebooklm.google.com/notebook/89ac8…
One subtle edge of MCP over A2A: MCP usage drives model training, which improves tool handling, creating a self-reinforcing loop. A2A lacks this flywheel effect currently.
Interesting tidbit from the kimi K2 paper They sampled 3000 real MCPs and used them to generate 20,000 synthetic tools Then generated thousands of agents with different prompts using those tools MCPs = please format your api exactly for our training run k thx
A Grok Heavy style group-chat where a bunch of these agents with different personalities debate and arrive at an answer would be useful and fun
The new "can select personality" feature is now visible in the ChatGPT web app - available personalities are Cynic, Robot, Listener, Sage, and Default Each personality appears to influence the system message prompt's personality section, including detailed instructions about…
GitHub's remote MCP Server returns 89 tools. OpenAI suggests fewer than 20 tools in your context for higher accuracy.
Lemonade stand posters these days aren’t merely better than before — they’re disturbingly professional
GitHub MCP server has a tool to assign an issue to Copilot Agent! So, you can deep-research and brainstorm the spec etc in the AI tool of your choice and then ask Copilot to implement it.
Kicking off async work via the GitHub coding agent has been a force multiplier for team. #github #ai #codingagent
Does the 5 mins meeting is actually 30 mins cost argument apply in the age of supervising coding agents? It seems to be actually more permissive of that -- kick off a coding task, check up on your msgs, chat with someone quickly... get back and review the code.
At first it looked like the future was for AI to review our work. Productivity was capped by how fast we could work. Now, it looks like the future is that we’ll send off tasks to AI Agents and orchestrate and review their work. Making productivity nearly uncapped.
Aaron Levie says the future of work has flipped. We thought AI would fix our mistakes. Now it makes the first draft. We QA what comes back. “The human's job is to fix the AI errors, and that's the new way we are going to work.”