🤖 OpenAI’s New Cloud-Based Software Engineering Agent

☁️OpenAI’s New Cloud-Based Software Engineering Agent

⚙️LLMs Get Lost In Multi-Turn Conversations

🛠Trending Tools

🥪Brief Bites

💰Funding Frontlines

💼Who’s Hiring?

It’s powered by codex-1, a tailored version of OpenAI o3 that’s optimized for complex software engineering workflows.

OpenAI o3 is a reasoning model designed to handle more complex multi-step problems. It leverages Chain-of-Thought (CoT) techniques to break down these more complex multi-step problems into manageable sub-problems. Then, it solves each manageable sub-problem, combining them into a complete solution.

Pro, Team, and Enterprise users can access “Codex” today through the sidebar within ChatGPT. To assign “Codex” a new coding task, you simply provide an input and click “Code.” If you also want to ask “Codex” a question about your codebase, you click “Ask.”

What makes “Codex” special is that it introduces parallelism: the ability to manage multiple parts of a complex software engineering workflow at the same time.

Just look at your web browser right now. How many tabs do you have open? Now, what’s happening in all those tabs? Each browser tab is static. You can’t have one tab open for writing code and another tab open for writing a Slack message at the same time.

Multi-turn conversations occur when user instructions are given in multiple stages rather than all at once. In other words, it’s a back-and-forth dialogue that unfolds over several inputs, or “turns.”

They examined OpenAI’s GPT-4.1, Anthropic’s Claude 3.7 Sonnet, and Google’s Gemini 2.5 Pro across six conversational tasks, analyzing over 200,000 simulated back-and-forth dialogues.

They discovered that each LLM’s accuracy and performance dropped by an average of 39% across all six conversational tasks when inputs were split over multiple turns.

In contrast, the same LLMs achieved a 90% success rate across all six conversational tasks when using single-turn conversations: when user instructions are given as a single, complete input or “prompt.”

Understanding that even the most capable LLMs significantly struggle with multi-turn conversations highlights the need for clear, concise, and cohesive inputs.

By providing a single, well-structured “prompt,” you reduce the chance of the LLM you’re using from getting “lost” in all the dialogue. So, aim to condense your “prompts” to maximize the effectiveness of the outputs you receive.

Somite raises over a $47M Series A to revolutionize cell therapy with AI.

Moonvalley lands a $53M Series B to craft high-definition generative videos with AI.

Granola secures a $43M Series B for an AI-based notepad that autonomously takes notes on your behalf.

Haize Labs (New York City, NYC): Software Engineer Intern, Fall 2025

NVIDIA (Santa Clara, CA): Firmware Engineer, Entry-Level

Recidiviz (New York City, NYC): Policy Data Analyst, Mid-Level

Postman (San Francisco, CA): Senior Data Analyst, Senior-Level