Jul 13, 2026

Deploying Google's Gemma-4 models on AWS Inferentia2 exposed fundamental incompatibilities: cross-layer KV-sharing, mixed query/KV head counts, and interleaved sliding-window/global attention break the AWS Neuron stack. The standard `optimum-neuron` library lacks a Gemma-4 model class, `neuronx-distributed`'s static graph builder cannot express KV-sharing, and the public Neuron vLLM endpoint output fluent but incorrect results. Engineers bypassed the abstraction to trace the Hugging Face forward pass directly. Meanwhile, a pricing audit revealed that tokenizer differences make $/token comparisons misleading: the same code file produced 681 tokens on one model's tokenizer and 1,178 on another's, and a same-vendor model update silently increased token counts by up to 39%, effectively a hidden price hike.

A week-long experiment wiring six MCP servers into a single agent logged 41,200 tool calls and showed that tool schemas alone consumed 41% of the context window, dropping prompt cache hit rates from 92% to 48% before any user input. Streamable HTTP added roughly 87% latency compared to local stdio, and rewriting vendor tool descriptions proved the cheapest optimization. In the broader agent landscape, a study found 38% of AI teams now cite evaluation as their top blocker: offline evals on fixed test sets measure the past, while real traffic drifts immediately, and multi-agent systems create exponential execution paths with emergent failures that single-agent evals miss.

New security disclosures highlight the weaknesses of AI-assisted code review. GhostCommit hid a secret-stealing payload inside a PNG image, with an AGENTS.md file innocently pointing to it. Because text-based reviewers like CodeRabbit exclude images by default, the attack slipped past both Cursor Bugbot and CodeRabbit without a single flag. Separately, a growing concern among engineers argues that AI-generated diffs break the practice of diff-based code review, as the artifact no longer contains the human reasoning traces senior reviewers used to infer intent.

In other security guidance, a pre-deploy checklist for Solana Anchor programs distills lessons from the 2022 Wormhole bridge exploit (a missing sysvar check enabled a $326M unauthorized mint) into systematic checks around account validation, signer verification, and safe PDA derivation. Claude Code's Skills feature evolved into a directory-based SKILL.md structure standardized at agentskills.io, designed to enforce conventions while reducing token overhead via dynamic context injection. Finally, a six-week audit of AI visibility dashboards found that blending branded and unbranded prompts inflates scores — stripping branded queries can turn a reported 35% visibility ranking to zero, exposing an artifact-driven illusion of discovery.