AI Builders
Digest

SAP CTO Philipp Herzig frames AI as a business model transition rather than just a technology shift. As the leader in end-to-end enterprise operating systems serving 400,000 customers, SAP is embedding AI agents, generative UIs, and reengineering data layers to deliver verifiable outcomes across finance, HR, supply chain, and more. Major challenges include scaling across complex landscapes, data fragmentation from M&A or legacy systems, and security—prompt injections or vulnerabilities like recent ones can expose credentials if not properly isolated.

Herzig highlights that LLMs excel in unstructured worlds (text, documents, support) but fall short for predictive/tabular tasks like demand forecasting or cash flow predictions, where classical ML or new architectures like RPT1 (relational pretrained transformers) are needed. Agent success hinges on evals for verifiability, reviving test-driven practices now that coding is automated. SAP's products like Joule for consultants demonstrate 30% effort reduction in complex migrations.

Quote: "The gap almost increases... versus getting narrow" between innovation and outcomes. SAP is moving to hybrid consumption-based pricing as trust and verifiability grow, while upleveling roles so humans focus on strategic work. Herzig also eyes quantum for hard optimization problems like logistics.

Anthropic's new Managed Agents is a hosted service for long-horizon agents via minimal, durable interfaces designed to outlast specific implementations. They virtualized three components: the session (append-only durable log of events, living outside Claude's context window), the harness (orchestration loop), and the sandbox (execution environment for code and file edits).

Previously, coupling everything in one container created "pet" servers that required nursing on failure. Now, the brain (Claude + harness) calls hands via a simple execute(name, input) → string interface. Containers become cattle—failures are handled as tool errors, new ones provisioned on demand. Harness failures are recovered by rebooting with the session log. Security boundaries ensure credentials never reach the sandbox where generated code runs.

Result: massive latency wins (p50 TTFT down ~60%, p95 >90%), easier VPC integration, support for many brains and many heterogeneous hands. The session enables flexible context interrogation without irreversible compaction. Quote from the post context: abstractions like process and file in OS outlasted hardware; Managed Agents aims for the same with agents.

Link: https://www.anthropic.com/engineering/managed-agents

Ryo Lu, Design at Cursor_ai, has fully switched to GPT-5.5 + Composer 2 in Cursor for the ideal balance of intelligence, speed, and cost. Try /multitask to break out of the queue and work on many things at once.

Josh Woodward, VP at Google Labs / Gemini / AI Studio, announced NotebookLM can now auto-label & categorize sources.

Replit CEO Amjad Masad shared: Import your Vercel or Lovable apps to Replit with a few clicks.

Peter Steinberger (Polyagentmorous ClawFather at OpenClaw + OpenAI) released discrawl 0.6.0 with the biggest new feature: reading Discord DMs without custom login tricks (no writing).

Box CEO Aaron Levie explains Jevons Paradox in the AI context: If AI makes employees more productive, companies will want as many productive units of labor as possible. As AI improves, it enables taking on more tasks, leading to hiring more people for surrounding work. Small businesses that couldn't afford complex software now can with AI-amplified engineers (5-10X impact). Sales teams with automated intelligence hire more reps for more leads; marketing with better video production hires editors, and so on. This trend will accelerate and appear in surprising ways.

OpenAI CEO Sam Altman announced GPT-5.5 and GPT-5.5 Pro are now available in the API. "This was a good week. Proud of the team. Happy building!"

Swyx highlighted Vercel's incredible AI engineering work presented at JSConf EU, coming directly from the C-suite—something that should be more common.

Dan Shipper, CEO at Every, noted: "Any AI knows more than any individual human. But any individual human learns faster than any AI."

Matt Turck noted that combining Cohere and Aleph Alpha highlights strong reluctance outside the US to outsourcing intelligence to a handful of US platforms amid geopolitical chaos.

Peter Yang demonstrated building Star Fox with GPT-5.5 and Codex after 15 min of prompting; the agent even play tests its own game for fun.