Market surveillance agent with LangGraph and Strands on AgentCore
AWS published a reference build pairing LangGraph orchestration with Strands agent reasoning on Bedrock AgentCore — a concrete pattern for compliance-grade multi-agent pipelines.
AI Daily Recap
20 articles · 5 categories
The finishable daily brief
What happened in AI — Jul 28, 2026
Tuesday, Jul 28, 2026
20 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Moonshot open-sourced Kimi K3, drawing distillation accusations against a U.S. model plus a report it's chasing more Nvidia GPUs for K4 despite export limits.
- AWS published a reference build combining LangGraph orchestration and Strands agent reasoning on Bedrock AgentCore for market surveillance.
- Google expanded Gemini API Managed Agents with a 3.6 Flash tier plus hooks and triggers.
- A formally verified 3D mesh-intersection implementation argues a 93-line Lean 4 spec beats 1,000 lines of AI-generated code.
- AWS launched a GuardDuty Investigation Agent that auto-triages security findings with MITRE ATT&CK mapping.
Moonshot open-sourced Kimi K3 today, and the release immediately turned geopolitical: a New York Times video pressed distillation accusations against an American model, while a separate report says Moonshot is seeking more Nvidia GPUs for Kimi K4 training in open defiance of U.S. export restrictions.
On the engineering side, production agent patterns kept converging — LangGraph-plus-Strands orchestration on Bedrock AgentCore, Gemini's expanded Managed Agents, disposable local sandboxes — even as coding-agent adoption ran into pushback over eroded team collaboration and a formally-verified counter-example to "just trust the agent."
Agent Runtimes & Orchestration in Production 4 items
Vendors and independent builders converged on the same shape for production agents today: an orchestration layer paired with isolated execution and an explicit, separable judgment component.
Gemini API Managed Agents: 3.6 Flash, hooks, and more
Google expanded Gemini API Managed Agents with a 3.6 Flash model tier plus hooks and triggers, pushing more of the agent lifecycle into the managed service rather than leaving it to the caller.
ALI: A Python agent architecture where the normative evaluator is
A new Python agent architecture keeps the 'normative evaluator' — the component that judges whether an action is acceptable — as a distinct, inspectable layer instead of folding judgment into the policy model.
Show HN: Hotcell – local sandboxes for AI agents
Hotcell is an open-source (Apache 2.0) tool for creating, pausing, and managing local agent sandboxes across devices, giving agents disposable, resumable execution environments without cloud infra.
Coding Agents: Adoption, Skepticism, and Practice 5 items
Coding agents keep scaling in production, but engineers are pushing back on what that costs in review discipline and trust, and reaching for formal verification or open-weight alternatives as a hedge.
Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI
OpenAI's product engineering lead detailed how Codex scaled to 10M users while building ChatGPT Work, describing Sites, OpenClaw, Memory, and Subagents as the actual architecture behind the assistant.
Scientific computing in the age of agentic AI
A new OpenAI field report documents scientists using AI coding agents to modernize scientific-computing codebases in genomics and other fields, citing concrete acceleration in development and discovery.
AI-coding agents kill team collaboration
A widely discussed critique argues AI coding agents erode team collaboration by letting engineers ship code without the discussion and review that used to force shared understanding.
Show HN: Formally verified 3D CSG: Trust 93 lines spec, not 1000 lines AI code
A developer published what they say is the first formally verified 3D mesh-intersection implementation, arguing a 93-line Lean 4 spec is more trustworthy than 1,000 lines of AI-generated code.
Show HN: Kdeps – A New CLI Coding Agent That Runs on Open Weight Models
Kdeps is a new open-source CLI coding agent built to run entirely on open-weight models, adding to the growing list of coding agents that don't depend on a frontier-lab API.
Kimi K3's Release and the Widening Open-Weight Fight 5 items
Moonshot open-sourced Kimi K3 today, and the release immediately spans model quality, IP disputes, and hardware-export politics rather than staying a simple model drop.
Moonshot Open-Sources Kimi K3 as U.S.-China AI Tensions Intensify - Caixin Global
Moonshot open-sourced Kimi K3 the same day U.S.-China AI tensions escalated, making the release itself part of the geopolitical story rather than just a model drop.
Video: Is Kimi K3 ‘Distilled’ From an American A.I. Model? - The New York Times
The New York Times examines claims that Kimi K3 was distilled from an American model, the latest entry in a recurring dispute over whether Chinese open-weight labs train on competitors' outputs.
Moonshot Openly Defies The Trump Administration By Seeking Access To Additional NVIDIA GPUs For Training The Next-Gen Kimi K4 Model - Wccftech
Moonshot is reportedly seeking additional Nvidia GPUs to train the next-generation Kimi K4, openly pushing against Trump-administration export restrictions rather than routing around them quietly.
Moonshot AI’s Kimi K3 Model: What We Know - Built In
A rundown of what's actually confirmed about Kimi K3 — Moonshot's newest model — for readers trying to separate specs from hype.
[AINews] Much ado about Open Weights
Latent Space's roundup notes that amid a flood of open-weights commentary, Kimi K3 was the only model that actually shipped today.
Enterprise AI Infrastructure & Data Stacks 4 items
Cloud and data vendors keep pushing retrieval and natural-language querying directly into the data layer — Postgres, BI, observability — instead of leaving it to a separate agent framework.
AI-Powered Retrieval in PostgreSQL with Azure HorizonDB
Microsoft detailed adding AI-powered retrieval directly into PostgreSQL via Azure HorizonDB, moving vector/semantic search closer to the database instead of a bolt-on service.
Bringing Conversational Analytics to your entire data ecosystem
Google Cloud expanded conversational analytics across its data ecosystem, arguing that trustworthy natural-language querying over business-critical databases needs governance built in, not a chatbot wrapper.
How LangChain Built an Agent-First Data Stack
LangChain describes combining Hex, dbt, semantic models, and observability into an 'agent-first' data stack that scaled its internal self-service analysis 40x.
Grafana Assistant Expands to More Than 30 Data Sources
Grafana Assistant now queries and correlates data across more than 30 sources in natural language, extending AI-powered observability beyond a single backend.
Security, Identity, and Automated Threat Response 2 items
Security teams are handing more identity management and incident triage to automated agents — federated machine identity at Best Buy, and AWS's new investigation agent for GuardDuty findings.
Best Buy scales AI workloads and secures access with Workforce Identity Federation
Best Buy adopted Workforce Identity Federation to manage access as it scaled AI workloads on Google Cloud, addressing the administrative and risk burden of syncing thousands of machine identities.
AWS Launches Amazon GuardDuty Investigation Agent to Automate Threat Triage
AWS's new GuardDuty Investigation Agent (public preview) auto-correlates findings, 90-day activity logs, and resource topology into structured reports with risk ratings and MITRE ATT&CK mapping.
You are caught up for this edition