🔥 Story of the Day
AWS’s Dogwood promises to fix complex AI agent tool call sequences — The New Stack
AWS introduced Dogwood, an open-source policy language designed to govern the sequence of AI agent tool calls, moving beyond single-action validation. Unlike simpler authorization models, Dogwood builds on concepts from Cedar by allowing policies to evaluate the historical context of the agent's interactions. This means you can enforce complex business rules that depend on the trajectory of the workflow—for instance, ensuring a refund tool call is only valid if a prior approval step, recorded in the agent's history, was successfully executed.
This is a critical architectural advance for reliable agentic workflows. Previously, governance tools could answer "Is this tool call valid right now?" Dogwood answers, "Is it permissible for the agent to call this tool after it has done X and before it does Y?" This ability to constrain multi-step, stateful decision-making is essential for moving AI agents into high-stakes, autonomous operations within regulated environments.
For those building ML infrastructure that relies on agents, this implies a need to implement stateful, sequence-aware authorization checks rather than relying on stateless API gateway guards. The technical shift is moving from simple perimeter security to deep workflow choreography control at the policy layer.
⚡ Quick Hits
Humans missed 1 in 3 threats approving AI agent commands across 40k game runs — Hacker News - Best
The article discusses tracking and analyzing the specific permissions consumed by AI agents across large simulation runs. The concrete insight is the visibility into which agents request which permission types. For Kubernetes deployments hosting LLM agents, this permits granular auditing and enforcement of the principle of least privilege, drastically tightening the operational attack surface.
Context Engineering in an LLM Harness — Hacker News - LLM
This approach structures context for LLMs using a formal ontology, moving beyond basic RAG retrieval. Instead of passing a list of text chunks, context is modeled as an ontological graph. This allows the system to explicitly reason about the defined relationships between different data sources, adding structural reliability to the context injection layer.
CloudflareOS: Why AI tools know nothing about your company — until now — The New Stack
CloudflareOS provides an open-source AI workspace layer designed to secure agent interaction with internal tools. It addresses context loss by controlling access to private enterprise systems, functioning as a secure boundary that prevents the leakage of proprietary data via raw API key distribution across multiple agent workflows.
Anthropic recommends a git worktree per agent. Your runtime infra makes that a problem. — The New Stack
Anthropic's recommendation for per-agent Git worktrees exposes that modern MLOps stacks must support parallel validation. The bottleneck lies in downstream systems (staging, queues) optimized for single, linear testing paths, which cannot handle the high merge rate of simultaneous agent iterations.
Does Kubernetes DRA Replace HAMi? — CNCF Blog
The GA of Dynamic Resource Allocation (DRA) in K8s v1.34 provides native scheduler support for requesting fractional device memory. While this simplifies resource definition compared to prior methods, the article notes that DRA does not cover the fine-grained, in-container enforcement of resource limits that specialized tools like HAMi previously managed.
Shadow AI in CI/CD: Threat-modeling the path from developer laptop to Kubernetes — CNCF Blog
"Shadow AI" refers to unvetted AI integrations in the SDLC. The risk profile changes when AI moves from advice to action. Infrastructure teams must treat every AI-executed action as an attacker's action, mandating that every agent identity must be registered, owned, constrained by least privilege, and monitored at runtime.
⚡ Quick Hits
SciCode-Verified: How Benchmark Defects Underestimated LLM Scientific-Coding — Hacker News - LLM
Content inaccessible.
Inside vLLM: Anatomy of a High-Throughput LLM Inference System (2025) — Hacker News - LLM
vLLM uses PagedAttention to optimize LLM inference by managing the KV cache efficiently, analogous to OS virtual memory paging. This directly boosts throughput, which is a primary lever for controlling the cost-efficiency of scaling self-hosted LLM inference endpoints on Kubernetes.
Researcher: gemma4:e4b • Writer: gemma4:e4b • Editor: gemma4:e4b