Issue #005 · September 20, 2026
This week, the interesting shift is happening around the agent, not inside the model.
As agents move from demos to real systems, the engineering challenges are becoming clearer: secure tool access, isolated execution, reliable evaluation, retrieval quality, and knowing whether faster code generation actually translates into shipped software.
This issue looks at that production layer — from real MCP vulnerabilities and agent security boundaries to sandbox infrastructure, evaluation economics, and new approaches to RAG.
GitHub Security Advisory
A critical vulnerability in @zereight/mcp-gitlab allowed an unauthenticated attacker to read arbitrary local files through an exposed MCP tool and potentially recover the GitLab access token used by the server.
CVE / project advisory
A separate vulnerability in the same GitLab MCP server allowed a malicious website to reach a locally running Streamable HTTP MCP endpoint through DNS rebinding.
CVE / GitHub Security Advisory
atomic-agents-stack accepted MCP server catalogs over cleartext HTTP. A network attacker able to modify the catalog response could inject commands that the agent host subsequently spawned as local processes.
Google Cloud
Agent Substrate is an open-source Kubernetes runtime designed specifically for high-density isolated agent execution, including fast sandbox suspend/resume and network/kernel isolation.
AgentsCamp
The observability market is now differentiated enough that “add tracing” isn't a useful architecture decision. Teams need to decide whether their primary requirement is debugging trajectories, release-gating with evals, OpenTelemetry integration or self-hosted data control.
NBER
A study combining AI-tool telemetry with data from more than 500,000 GitHub developers examines whether productivity improvements from autocomplete, interactive agents and autonomous coding agents survive all the way to shipped software.
Academic research / arXiv
An empirical study asks when the additional ingestion cost of graph-based RAG actually produces enough retrieval quality to justify building the graph
Independent open-source project
Ouroboros experiments with persistent identity and durable memory across tasks while allowing the agent to modify its own implementation, prompts, tools and dependencies
Get the next one straight to your inbox.