Claude Code Adds AGENTS.md Support as a CLAUDE.md Fallback
Claude Code v2.1.277 now falls back to AGENTS.md when no CLAUDE.md is present, folding a competing convention into its own config discovery.
AI Daily Recap
16 articles · 6 categories
The finishable daily brief
What happened in AI — Sep 18, 2026
Friday, Sep 18, 2026
16 articles · 6 categories
read top to bottom · then stop
In 30 seconds
- Claude Code (v2.1.277) now falls back to AGENTS.md when no CLAUDE.md is present, folding a rival convention into its own config discovery.
- DoorDash's multi-agent LLM system retired 60,000 stale feature flags across 623 repositories, with engineer approval gates in the loop.
- Anthropic and Accenture will invest more than $1 billion combined to build independent evaluation capacity for frontier AI.
- Zhipu's GLM-5.3-FlashX runs near 200 tokens/sec on roughly 100,000 domestically made accelerators.
- A new report says OpenAI and Anthropic's Chinese rivals still capture only a fraction of their revenue despite lower operating costs.
Coding-agent tooling kept consolidating and scaling up today: Claude Code added AGENTS.md as a fallback convention, and production teams showed agents doing large operational work rather than just writing code — DoorDash's multi-agent system retired 60,000 stale feature flags across 623 repos, and Google detailed using agentic AI to help secure hundreds of millions of lines of its infrastructure code.
Evaluation is getting institutional backing to match: Anthropic and Accenture committed over $1 billion combined to build independent evaluation capacity, and OpenAI published an internal triage framework for reporting model misalignment. Chinese labs kept shipping — Zhipu's GLM-5.3-FlashX and Moonshot's climb toward a $50 billion valuation — but a new report says their revenue still trails OpenAI and Anthropic despite lower costs.
Coding-Agent Standards Consolidate Around AGENTS.md and MCP 3 items
Three signals point the same direction: coding-agent tooling is consolidating around shared conventions instead of one-off formats — Claude Code adopting AGENTS.md, a dedicated fast-decision model slotting into the agent loop, and public debate over whether Skills has already superseded MCP and RAG.
Should You Read the Code, Is RAG Dead, and Did Skills Kill MCP?
GitHub's podcast crew debate whether Anthropic's Skills format has displaced MCP and whether RAG still earns its keep against long-context coding agents.
What Is Jev? TypeSafe AI's Fast "System One" Model for Agent Loops
TypeSafe AI's Jev is a small "System One" model built for fast, structured decisions inside a LangChain agent loop, distinct from the LLM doing the slow reasoning.
Agents Move From Coding Assistants to Large-Scale Ops Automation 3 items
Today's deployment stories move past coding assistance into large-scale operational work: DoorDash's agents retired 60,000 feature flags, Google secured hundreds of millions of lines of infrastructure code, and AWS packaged agent skills to handle model deployment end to end.
DoorDash Uses Multi-Agent LLMs to Clean Up 60,000 Feature Flags
DoorDash's multi-agent LLM system retired 60,000 stale feature flags across 623 repositories, combining live experiment data via MCP with engineer approval gates.
How Google Uses Agentic AI to Secure Hundreds of Millions of Lines of Code
Google detailed how it uses agentic AI to scan and secure hundreds of millions of lines of infrastructure code against emerging AI-assisted exploit techniques.
AWS Ships Six Coding-Agent Skills to Deploy Hugging Face Models on SageMaker
AWS packaged six open-source agent skills that let a coding agent deploy Hugging Face models on SageMaker AI, auto-selecting serving containers and autoscaling.
AI Infrastructure and Inference Scaling 2 items
AWS's new GPU-aware inference router and vLLM's use of NVIDIA's hardware video decoders both squeeze more inference throughput out of existing GPUs rather than just adding more of them.
Amazon SageMaker HyperPod Inference Gateway Routes by Live GPU Load
SageMaker HyperPod's new Inference Gateway is a Kubernetes-native, GPU-aware router on EKS that uses live GPU signals to cut first-token latency.
vLLM Scales Multi-GPU Video Captioning With NVIDIA's Video Decoders
vLLM detailed using NVIDIA's hardware video decoders to scale multi-GPU video captioning and description workloads.
Open-Weight Releases From Chinese Labs 2 items
Zhipu pushed a fast open-weight model on domestic accelerators while DeepSeek's newly open-sourced execution harness reframes what counts as "industrial-grade" agent engineering.
Zhipu's GLM-5.3-FlashX Hits Near 200 Tokens/s on Domestic Chips
Zhipu's GLM-5.3-FlashX runs near 200 tokens/sec on a cluster of roughly 100,000 domestically made accelerators, its latest bet on chip self-sufficiency.
DeepSeek's Harness Goes Open Source, Sparking Debate on Agent-Engineering Boundaries
36Kr examines DeepSeek's newly open-sourced execution harness and the industrial-grade boundaries it sets for production agent engineering.
Evals and Safety Get Institutional Backing 3 items
Independent evaluation of frontier AI is getting real money and process behind it — Anthropic and Accenture's $1 billion-plus commitment, OpenAI's internal misalignment-reporting framework, and a new open benchmark for agent memory.
Anthropic and Accenture Commit Over $1 Billion to Independent AI Evaluation
Anthropic and Accenture will invest more than $1 billion combined to build independent, embedded evaluation capacity for frontier AI models.
OpenAI Introduces a Triage Framework for Reporting Model Misalignment
OpenAI published a triage framework letting employees flag suspected model misalignment, plus initial case studies of unexpected model behavior.
An Open Benchmark Aims to Standardize How Agent Memory Systems Are Compared
A new open leaderboard standardizes Add/Search benchmarking for agent memory systems, aiming to stop each vendor from grading its own homework.
Chinese AI Labs' Revenue Gap Widens as Compute Scales 3 items
Even as Moonshot's valuation and Z.ai's compute position keep climbing, a fresh report says Chinese model providers are still capturing only a fraction of OpenAI and Anthropic's revenue.
OpenAI and Anthropic's Chinese Rivals Face a Revenue Reality Check
A new report says OpenAI and Anthropic's Chinese rivals are still capturing only a fraction of their revenue despite running at lower cost.
Inside the $50 Billion Rise of Moonshot AI
ThinkChina traces Moonshot AI's climb to a $50 billion valuation and the bets behind it.
Z.ai's Next Growth Engine Starts Where Compute Scarcity Ends
Digitimes argues Z.ai's next growth phase begins as domestic compute scarcity eases, shifting the constraint from chips to product.
You are caught up for this edition