Issue #002 · August 30, 2026
This week, the interesting shift isn’t another model release. It’s what happens around the model: how agents are contained, evaluated, given memory, connected to tools, and governed once they start operating real systems.
That’s where agent engineering starts to look a lot more like systems engineering.
GitHub
OpenClaw's maintainers share what happens when AI-generated contributions scale faster than the human review systems around them — from “prompt requests” to new trust and security signals.
Trail of Bits
Trail of Bits gave a cyber-capable agent a VM-escape challenge. It used known vulnerabilities, combined bugs that were not marked as security issues, and eventually constructed an exploit chain involving several zero-days. Trail of Bits tested whether a modern cyber-capable agent could escape a conventional QEMU/KVM sandbox. It succeeded through multiple paths, eventually chaining newly discovered vulnerabilities.
Pillar Security
Pillar describes a vulnerability chain where the changed threat model around an agentic GitHub workflow exposed credentials and enabled compromise of a sandbox GCP project. Google subsequently tightened tool scopes and credential access.
IBM Research / Hugging Face
IBM Research evaluates memory strategies across eight models and finds that more memory isn't automatically better. Some models benefit from the full accumulated guideline set; others perform better with selective retrieval.
PulseMCP / O’Reilly
The article moves beyond single MCP-server demos and looks at compositions where multiple clients and servers work across applications and workflows. The real value of MCP starts when workflows cross application boundaries. At that point, configuration, authentication, tool discovery and governance become much more interesting than simply exposing one API to one agent.
DeepSeek / GitHub
DeepSeek Harness is an MIT-licensed agent harness built around a plugin-based architecture where runtime capabilities can be composed and replaced. It is explicitly still a developer preview. The harness layer is becoming as important as the underlying model: tools, state, sessions, permissions and execution loops increasingly live outside the LLM itself.
arXiv
MobileJudgeBench evaluates LLM-as-judge systems across 931 human-annotated agent trajectories, six mobile-agent benchmarks and 68 applications. One notable result: a relatively simple judge can compete with more elaborate pipelines.
SANS
This one is excellent. 90 practicing incident responders ran 1,775 evaluations against 123 agent harnesses, including adversarial testing. The top five implementations are open source. Most agent benchmarks are constructed by AI researchers. Here, actual incident responders tested whether autonomous agents could investigate compromised systems while resisting destructive or misleading behavior.
Datadog
Datadog partage ce qu’ils ont appris en observant leurs propres agents internes : les logs applicatifs classiques montrent souvent l’action finale, mais pas le prompt, le contenu récupéré ou le résultat d’un tool qui a réellement conduit l’agent à cette action.
Get the next one straight to your inbox.