Architecture AI Updates: October 18, 2026

1. Eight Chunking Strategies for RAG-Based Agents, From Fixed-Size to Late Chunking

ByteByteGo. ByteByteGo’s EP229 newsletter lays out eight ways to split documents for agent retrieval: fixed-size with overlap, recursive (paragraphs, then lines, words and characters), document-based on headings or tags, semantic (break where embedding similarity between neighboring sentences drops), LLM-based rewriting into standalone statements, agentic (an agent picks a method per document type), late chunking, and hierarchical layers of sections, paragraphs and sentences. Late chunking embeds the whole document first and then averages token embeddings within each section, which requires the document to fit in the embedding model’s context window. LLM-based chunking gives high quality but is expensive across large corpora. Source

2. “AI Security Is Still Kubernetes Security”: Isolating GPUs, Models and Agents

InfoQ (Andrew Martin). In a QCon London talk published by InfoQ, controlplane.io CEO Andrew Martin argues that most AI platforms run on Kubernetes, so cluster hardening remains the base layer, with GPU clusters treated as crown-jewel assets. He advises against GPU time-slicing for sensitive multi-tenant data in favor of hardware partitioning, dedicated GPUs or confidential computing, and recommends separating workloads by data classification. For agents, he recommends treating them as untrusted actors, testing outputs against ranges and bounds rather than fixed contracts, running coding agents in sandboxes, and having MCP servers use delegated user identity with user-scoped RBAC and read-only modes. Model provenance and signing (SLSA for AI) and a prompt firewall are presented as additional layers, with the caveat that input filtering alone is not reliable. Source

3. Reverse Engineer Anything Exposes Ghidra, Hopper and IDA to Coding Agents Over MCP

InfoQ (Renato Losio). InfoQ covers Reverse Engineer Anything (REA), an MIT-licensed project that connects coding agents such as Claude Code and Codex to disassemblers through MCP servers. The agent queries Ghidra, Hopper or IDA, inspects functions and cross-references, and tests hypotheses against the binary in an iterative loop; JavaScript, Electron and .NET targets can be analyzed without a native engine. Cited results include reconstructing a DX-Ball sound-panning function that matched all 3,205 original x86 test cases and all 63 compiled bytes. Hacker News commenters questioned what REA adds over asking an agent to drive Ghidra directly, and developers raised concerns about copying games and bypassing paid features. Source