Planet AI Weekly: Planet AI Weekly May 17, 2026

This week we reviewed 224 articles from 14 sources. Here is what mattered.

Agentic tooling moved from demos into production infrastructure as Codex expanded across sales, engineering, and enterprise workflows at Sea and Databricks. The Musk v. Altman trial wrapped and given to the advisory jury while new evaluation layers, agent runtimes, and cross-account data controls tackled the gap between benchmarks and reliable deployment.

Official Highlights

OpenAI documents how sales teams use Codex to generate pipeline briefs, forecast reviews, and stalled-deal diagnoses from real CRM inputs. Worth a look for the prompt structures and output schemas if you run revenue operations. Read more

AWS ships Claude Platform on AWS, routing Anthropic's native Claude environment through existing AWS accounts and billing. Teams already on Bedrock can test production agent workflows without onboarding a new vendor. Read more

Hugging Face and AWS map the building blocks for foundation model training and inference on AWS, including data pipelines and distributed setups. Useful reference if you are scaling past single-node experiments. Read more

OpenAI profiles Databricks running GPT-5.5 in enterprise agent workflows after the model set a new state of the art on OfficeQA Pro. Worth the click for the deployment shape, not the benchmark line. Read more

Jensen Huang told Carnegie Mellon graduates the industry is rebuilding around AI-first infrastructure, with specific guidance on where to point technical skills next. Read more

From the Community

Nous Research proposes Lighthouse Attention, a training-only hierarchical selection mechanism that delivers 1.4-1.7x pretraining speedup at long context and gets removed before inference. Read more

Vercel Labs released Zero, a systems language that emits JSON compiler diagnostics so AI agents can read errors, repair code, and ship native programs without human interpretation. Read more

LangChain ships Delta Channels in LangGraph 1.2, a checkpointing primitive that stores only state diffs each step and writes periodic snapshots, killing the O(N squared) storage growth that breaks long-running agents. Read more

BerriAI open-sourced LiteLLM Agent Platform, a Kubernetes-based layer for isolated agent sandboxes and persistent session state across restarts and teams. Read more

Cline extracted its agent harness into an open-source TypeScript SDK that now powers its CLI and Kanban, with VS Code and JetBrains extensions being migrated onto the same four-layer runtime. Read more

Zyphra converted an autoregressive MoE into a discrete diffusion model, releasing ZAYA1-8B-Diffusion-Preview with up to 7.7x inference speedup and no systematic loss on evals. Read more

Fastino Labs open-sourced GLiGuard, a 300M parameter moderation model that handles prompt safety, jailbreak detection, harm classification, and refusal detection in a single forward pass, matching models 23-90x its size. Read more

Featured This Week

Emmimal P Alexander argues most LLM evals are vibes dressed up as metrics, then shows the receipts: a lightweight Python layer that forces every model output through explicit attribution, specificity, and decision-threshold checks before anything ships. The post pairs runnable code with the failure modes any team running agents or copilots in production will recognize - hallucinated citations, overconfident vague answers, judge models marking their own homework. Copy the thresholds, swap in your own rubric, and you have a reproducible gate between experiment and release. The clearest piece of the week on closing the gap between benchmark scores and shipping decisions. Read more