OpenAI president claims AI now writes 80% of your code

TLDR: OpenAI says agents now write most of your code, AWS ships a serious LLM migration playbook, and Sun Finance posts eye-popping IDV numbers on Bedrock.

OpenAI president: agentic tools now write “80% of your code”

In a Sequoia Capital talk, OpenAI president Greg Brockman said agentic coding tools jumped from writing 20% to 80% of software engineers’ code between December and now, as of 2026-05-01. He framed this shift as moving AI from a “sideshow” to the main act in software development, implying that models are now orchestrating multi-step coding tasks instead of just autocomplete.

If those numbers hold up in real workflows, this is a signal that agent-based IDE copilots and CLI agents are close to default for greenfield code. For engineering leaders, that means revisiting hiring profiles, coding guidelines, and review practices to handle machine-generated code at scale. The big missing piece is detail: no repos, benchmarks, or methodology yet, so treat the 80% as directional, not a hard metric.

Read more →

---

AWS publishes framework for LLM migration in production

Amazon Web Services released the “AWS Generative AI Model Agility Solution,” a framework for migrating and upgrading large language models (LLMs) in production environments, as of 2026-05-01. The guide walks through prompt conversion, evaluation, rollout, and tooling patterns so you can swap models without breaking downstream apps.

This is aimed squarely at teams stuck on “v1 forever” because model changes are scary to deploy. It covers prompt normalization, test harnesses, and operational playbooks so you can move between Amazon Bedrock, self-hosted models, or third-party APIs with less risk. The post is light on hard benchmarks, but rich in architecture diagrams and concrete mechanisms.

If you maintain multi-tenant generative AI services or internal platforms, this is worth a careful read; it is essentially AWS’s opinionated spec for model lifecycle management in enterprises.

Read more →

---

Sun Finance shares real numbers on ID verification with Bedrock

Sun Finance detailed how it built an AI-powered identity verification and fraud detection pipeline on Amazon Web Services using Amazon Bedrock, Amazon Textract, and Amazon Rekognition. The company reports extraction accuracy improvements from 79.7% to 90.8%, per-document cost reductions of 91%, and latency drops from up to 20 hours to under 5 seconds, as of 2026-05-01.

The architecture uses Textract for optical character recognition (OCR), Rekognition for image checks, and a large language model from Amazon Bedrock to structure and validate the extracted data. That hybrid pattern is notable: specialized vision plus a general LLM outperformed either alone. For anyone building know-your-customer (KYC), fintech onboarding, or document-heavy workflows, this is a solid reference design with concrete metrics instead of aspirational claims.

The system is also fully serverless, which matters if you want to scale up and down with spiky verification traffic without running a dedicated cluster.

Read more →

---

Quick Hits

• Exclusive: Citi moves into agentic AI Citi is rolling out Arc, a platform that connects multiple autonomous agents and use cases in one place so managers can monitor behavior and intervene, starting with internal developers.

• Unleashing Agentic AI Analytics on Amazon SageMaker with Amazon Athena and Amazon Quick Amazon Web Services shows how to wire an agentic analytics assistant that sits on Amazon Simple Storage Service (S3) data via Amazon SageMaker, AWS Glue, and Amazon Athena, then surfaces in Amazon QuickSight.

• Harvard AI Outperforms Doctors in ER Triage Study A Science paper reported by Gizmodo finds OpenAI o1-preview beat two attending physicians on 76 emergency department triage cases, with 67.1% correct calls vs 55.3% and 50.0%, raising sharp questions about deployment and liability.

• Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell A tiny agent harness written in portable shell with just sh, curl, and awk targets “no dependencies” environments and shows how far you can push system primitives, though maintainability will be a concern.

• Our evaluation of OpenAI's GPT-5.5 cyber capabilities Simon Willison highlights that the United Kingdom AI Security Institute found GPT-5.5 comparable to Anthropic Claude Mythos for vulnerability discovery, but unlike Mythos it is generally available now.

• Codex CLI 0.128.0 adds /goal The OpenAI Codex CLI coding agent now supports a /goal loop similar to the Ralph pattern, repeatedly planning and executing until it judges the goal complete or hits a token budget.

• Enabling a new model for healthcare with AI co-clinician Google DeepMind outlines an “AI co-clinician” vision to address clinician shortages, positioning models as tightly supervised teammates rather than replacements and hinting at longer term regulatory and workflow integration work.

• Scaling Pain of Coding Agent Serving: Lessons from Debugging GLM-5 at Scale Zhipu AI shares war stories from running coding agents on GLM-5 at scale, which is useful for anyone hitting strange failure modes in long-running tool-calling agents.

• Introducing Advanced Account Security OpenAI adds phishing-resistant login options, stronger recovery, and new data protection features that you should probably enable if your org relies heavily on their APIs.

• Quoting Andrew Kelley Andrew Kelley argues that LLM-assisted pull requests have a “digital smell” that experienced reviewers can spot, a reminder that human and AI coding styles still differ in detectable ways.