Wednesday, September 23, 2026
Claude Code silently drops local AGENTS.md instructions when privacy or enterprise proxy settings disable telemetry, requiring manual import workarounds.
In this issue · 6
Anthropic released Claude Opus 5.5, delivering Claude Fable 5.1 capability levels at 40% lower operational cost. Input tokens now cost $4 per million, while prompt-cache reads drop 60% to $0.20 per million.
NVIDIA released architectural mitigations in TensorRT-LLM for confidential computing on DGX B200 systems. It preserves over 96 percent throughput on DeepSeek-R1 while keeping token latency overhead below five percent.
The dart_mcp package introduced bounded pagination helpers capped at 64 pages to prevent infinite loops from buggy or malicious Model Context Protocol servers. It also added automated weekly conformance test suites and streamable HTTP examples.
Dynaris detailed an architecture for multi-location agent systems that prevents one workspace from touching another's data by isolating Model Context Protocol tool servers per workspace. The design unifies voice, chat, and email on one customer thread backed by session checkpointing, graph caching, and persistent workflow state.
NVIDIA published Nemotron 3 Diarization, an open-weight 100M-parameter model that separates up to eight speakers during overlapping speech. It hits a 14.72% diarization error rate on VoiceArena benchmarks and delivers streaming inference down to 320 milliseconds of input buffer context.
Claude Code version 2.1.277 introduced support for root AGENTS.md files, but the loader depends on a remote feature flag that fails when telemetry is disabled. Sessions omit project instructions without throwing warnings on offline shells, Amazon Bedrock, or Google Vertex AI. Developers can bypass the flag using an explicit import in CLAUDE.md.
Email digest
One email a day — the stories that matter for engineers, founders and tech leads. Human-edited, with links to primary sources.
By subscribing you agree to the privacy policy.