AWS ships Bedrock text-to-SQL blueprint for production apps

TLDR: AWS shipped a Bedrock text-to-SQL blueprint and cost controls, plus there is an interesting Hybrid Attention trick for faster inference.

AWS publishes Bedrock-based text-to-SQL reference pattern

Amazon Web Services has published a detailed guide for building a text-to-SQL workflow on Amazon Bedrock that turns natural language questions into executable queries and surfaces answers, as of 2026-04-08. The post covers orchestrating large language models with Amazon Bedrock to parse questions, generate SQL over your schemas, run queries, and return human friendly results.

For teams trying to productionize natural language analytics, this is essentially an opinionated reference architecture on AWS. You get patterns for handling schema context, prompt design, and routing queries, rather than just a toy notebook. The tradeoff is obvious: you are locked into Bedrock, and the post does not ship benchmarks or latency numbers.

If you are already on AWS and being asked to “let the execs talk to the data,” this is a good starting point to crib from and harden with your own guardrails.

Read more →

---

Amazon Bedrock Projects adds workload level AI cost tracking

Amazon Web Services introduced Amazon Bedrock Projects to attribute model inference costs to specific workloads and analyze them in AWS Cost Explorer and AWS Data Exports, as of 2026-04-08. The guide walks through end to end setup: defining Projects, designing a tagging strategy, and wiring that into existing AWS cost management.

If you run multiple agents or applications on Bedrock, this finally gives you first class cost visibility per project without building your own usage ledger. You still need discipline around tagging and deployment hygiene; mis tagged resources will quietly skew your numbers. There is also no magic new pricing, just better attribution and reporting.

For teams under pressure to show unit economics for AI features, this is worth a close look and probably a quick spike in a non critical environment to validate the tagging model.

Read more →

---

Hybrid Attention experiment shows big speedup with small perplexity hit

An independent researcher on Hacker News shared a Hybrid Attention modification to PyTorch and Triton that makes transformer attention linear in the first and last layers and quadratic only in the middle, as of 2026-04-08. On a Rust focused language model trained from scratch, reported inference improved from 17.96 seconds at 5.6 tokens per second with full attention to 0.35 seconds at 286.6 tokens per second, with a low perplexity increase on their tests.

For anyone pushing long context or real time agents, this is an interesting direction: change the attention pattern instead of only relying on kernel tricks. The caveats are big: this is a personal fork, evaluated on a single byte level Rust corpus, with no full paper or broad benchmarks yet. It is closer to a research note than a drop in production primitive.

If you maintain custom models or serve niche domains, Hybrid Attention is a good concept to track and maybe reproduce in your own sandbox rather than ship directly.

Read more →

---

Quick Hits

• Show HN: Frontend-VisualQA: give coding agents eyes to verify their own UI work Coding agents can call this CLI and Model Context Protocol (MCP) server to visually verify a web page against natural language claims, using Playwright under the hood, so they catch clipped modals and layout issues instead of only passing DOM checks.

• Show HN: Marimo pair – Reactive Python notebooks as environments for agents marimo pair lets AI agents operate inside a live Marimo notebook so they can use it as working memory and a reactive Python runtime, which is handy for data work and computational research where humans and agents iterate together.

• Artifact's Omni AI platform uses plain language to automate workflows Artifact parses natural language descriptions into multi step workflow graphs, then uses a multi agent architecture to construct and connect nodes for tasks like reconciliations and exception routing, targeting accounting and finance teams that want agentic automation without direct scripting.

• Deep Agents v0.5 LangChain’s Deep Agents 0.5 adds async subagents so remote workers can run in the background, plus better multimodal filesystem support, which should help if your agentic workloads need concurrent tool calls or heavy I/O.

• Arcade.dev tools now in LangSmith Fleet LangSmith Fleet can now route through Arcade, an MCP style runtime that exposes more than 7,500 agent optimized tools via a single secure gateway, giving production agents a curated tool universe with authorization and governance controls baked in.

• Alice, Lovable partner to test AI coding systems for security flaws Lovable is partnering with Alice for adversarial testing of its natural language app builder, using structured misuse scenarios to probe security holes in systems that have already seen more than 25 million user created projects.

• Mastercard to roll out authenticated agentic transactions in ASEAN Mastercard Agent Pay pilots in Southeast Asia combine tokenization with explicit intent verification so AI agents can initiate payments while merchants and consumers still get strong authentication and traceable authorization flows.

• GLM-5.1: Towards Long-Horizon Tasks Z.ai’s GLM 5.1 is a 754 billion parameter, 1.51 terabyte MIT licensed model available on Hugging Face and via OpenRouter; Simon Willison highlights its long horizon task focus and practical usage via his llm CLI.

• Anthropic's Project Glasswing: restricting Claude Mythos Anthropic is keeping its Claude Mythos model limited to vetted security researchers under Project Glasswing, which Simon Willison argues is a reasonable middle ground for a powerful system card flagged model.

• SQLite WAL Mode Across Docker Containers Sharing a Volume An experiment shows SQLite write ahead logging mode works reliably across two Docker containers sharing a volume on the same host, which is useful if your lightweight agent infra fans out workers over a shared SQLite store.