← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-04-23

2026-04-23 · Historical edition

X / Twitter

Claude, Anthropic

Memory in Claude Managed Agents has entered public beta across web, desktop, and mobile for all subscription plans. Agents can now learn from each session through an intelligently optimized memory layer balancing performance and flexibility. Memories are stored as files that developers can export and manage through the API, retaining full control over what agents keep.

Sam Altman, OpenAI CEO

OpenAI partnered with NVIDIA to try deploying Codex company-wide, with good results. Sam invited other companies interested in a similar rollout to contact them.

Aaron Levie, Box CEO

GPT-5.5 has launched. Box tested it over the past two weeks on its most complex knowledge-work evaluations and found enterprise-content accuracy improved by 10 percentage points over GPT-5.4. Specific results: financial services rose from 64% to 83%, healthcare from 61% to 78%, public sector from 59% to 72%, and media and entertainment from 57% to 70%. GPT-5.5 significantly improves advanced reasoning, data analysis, and handling complex context, and is now available to all customers in Box AI Studio.

Guillermo Rauch, Vercel CEO

Guillermo published a detailed update on attacks against Vercel. The team analyzed nearly 1 petabyte of network and API logs, extending beyond the original Context.ai breach. It found much broader activity: threat intelligence showed attackers distributing malware across multiple startups, specifically seeking Vercel accounts and other cloud providers’ tokens and API keys. Logs showed rapid enumeration of nonsensitive environment variables after obtaining keys. Vercel deepened collaboration with Microsoft, AWS, and Wiz and proactively notified other affected victims, recommending credential rotation and security best practices.

Amjad Masad, Replit CEO

Replit Agent can now be called directly from Gemini Enterprise. Replit’s Security Agent is also rolling out for app security reviews. Amjad released a white paper showing that combining existing LLMs with static-analysis tools can deliver performance improvements above 90% in some scenarios, even without frontier models such as Mythos.

Garry Tan, Y Combinator CEO

Garry shared a Claude Code skill-design lesson: when two skills have similar functionality, merge them into one larger skill with parameterized branches rather than retain separate skills. Fewer, larger skills shorten the resolver, reducing context consumption and improving performance.

Nikunj Kothari, FPV Ventures Partner

Nikunj shared what he currently considers the strongest personal stack: Claude Opus for planning and frontend design, OpenAI Codex for engineering, Conductor for orchestration, and Railway for deployment. Building has never been easier or more enjoyable, he believes. He also analyzed the M&A boom: worsening seed-to-Series-A success rates, liquidation of 2021 zombiecorns, extensive hiring by large-model factories, a weak IPO market, and flatter organizations prompting more people to found startups and widening the seed-to-A gap. He summarized with a quote: “There are many founders, but few real entrepreneurs.”

Dan Shipper, Every CEO

Dan’s GPT-5.5 vibe check found that many models write a good plan but then retreat from executing it. GPT-5.5 simply does the work.

Cat Wu, Anthropic Claude Code Engineer

Cat spoke with Lenny Rachitsky about sustaining Claude Code’s iteration speed, the evolution of product management in the AI era, and the future of work.

Podcasts

No Priors — SAP: Bringing the 'Operating System' of a Company into the AI Era with CTO Philipp Herzig

Key takeaway: SAP CTO Philipp Herzig sees AI reshaping enterprise software on three levels simultaneously. LLMs themselves are unsuitable for predictive tasks; the real challenge is not introducing AI but making it reliable at a trillion-scale volume of data.

Philipp Herzig is SAP’s CTO. SAP serves 400,000 enterprise customers across finance, HR, supply chains, manufacturing, logistics, procurement, sales, and other end-to-end processes. As the world’s largest enterprise software provider, he compares it to a company’s operating system.

Why SAP Has Survived Multiple Technology Cycles

Founded in 1972, SAP has crossed each technological wave from mainframes to client-server, internet, mobile, and AI. Philipp attributes its durability not to getting one era right, but to customers consistently needing outcomes and ROI. Software forms change, but underlying needs in order-to-cash and source-to-pay remain. He compares AI migration with the on-premises-to-cloud transition: people initially thought it meant moving software online, then realized CI/CD, multitenancy, and elasticity required fundamental rewrites. The same is happening with AI.

AI Is Reshaping SAP on Three Levels

The first is UI. Philipp believes the era of software requiring human intelligence sitting at a screen clicking interfaces to complete tasks has ended. SAP is pursuing generative UI, dynamically creating interfaces. For supply-chain analysis, users can describe a problem such as “What do new tariffs mean for my supply chain?” and receive deep analysis of SAP data rather than general retrieval like Perplexity. Systems can operate overnight and proactively report anomalies with recommendations in the morning, previously impossible proactivity. Second are business processes. Formerly rigid standard procedures can now combine structured and unstructured data through agents, turning “software as a service” into “service as a software,” with outcomes as the service. Third is data. SAP holds extensive structured core-business information—general ledgers, invoices, inventory—and is using SAP Knowledge Graph to bridge natural language and structured data through a unified semantic view that AI can navigate.

Scale Is the Real Engineering Challenge

Philipp said a RAG demo on 10 documents can impress a CEO, but 1000 documents pose real engineering challenges, and 10,000 APIs—SAP has 20,000—explode MCP context. Building AI is not hard; making it do the right thing at enterprise scale is. He particularly emphasized evals: agentic coding works fundamentally because output is verifiable through compilation and passing unit tests. Enterprise AI likewise needs evaluation sets built from historical system data—given this input, expect this output—to make coding agents reliable in business scenarios. He likened this to test-driven development’s return: previously few really practiced it, but now it is mandatory.

LLMs Are Unsuitable for Prediction; Traditional ML Is Needed

LLMs excel at unstructured text and knowledge work, but demand and cash-flow forecasting and receivables classification are fundamentally regression and classification problems. Traditional ML such as XGBoost and AutoML suits these tasks better, but these methods do not suit enterprise scale. SAP is building proprietary predictive and tabular models directly into workflows rather than making users improvise forecasts through natural-language tool calls in chat. He considers combining structured and unstructured data the hardest and most valuable part of enterprise AI implementation today.

Agent Mining and the Data Feedback Loop

Philipp introduced “agent mining,” analogous to process mining. When agents ask users for clarification, those decision traces and inputs become new training data and eval sources. Identifying unusual decisions and distilling good practices lets enterprises spread one regional team’s strengths globally, creating a data feedback loop that continually refines agents’ behavioral boundaries.