Patterns and problems in emerging multiagent systems (opens in a new tab)
Anthropic
Anthropic explores what happens when multiple frontier agents interact in shared environments, including cooperation, competition and unexpected systemic behaviors.
Issue #001 · August 23, 2026
AI DevList #001 — August 23, 2026
Multi-agent systems, MCP, agent skills, observability and production AI.
A focused selection of what mattered this week for engineers building AI systems.
Agentic AI
Introducing Dynamic Subagents in Deep Agents (opens in a new tab)
LangChain
LangChain explores generating small orchestration programs that coordinate subagents instead of relying exclusively on sequential model tool calls. The approach targets context isolation, parallel execution, branching, and more efficient orchestration.
AI Security & Red Teaming
Deadbugz: Currently Active MCP Supply-Chain Campaign — Pillar Security (opens in a new tab)
Pillar Security
Pillar Security uncovered an active MCP supply-chain campaign where a seemingly harmless server changes its tool metadata after a few normal calls, then attempts to steer the connected agent toward sensitive credentials and local configuration.
The State of MCP Security 2026 — AgentAuditKit (opens in a new tab)
AgentAuditKit / independent open-source research
An open-source security study scanned 571 real public MCP configurations and found recurring weaknesses around authentication, package pinning, secret exposure, shell execution and supply-chain risk.
LLM Engineering
State of Open Models: Summer 2026 (opens in a new tab)
Hugging Face
Hugging Face examines open-model activity across the first seven months of 2026, including releases, downloads, licensing patterns, derivatives and the increasing role of agents on the Hub.
Tools & Frameworks
The next generation of MCP (opens in a new tab)
Cloudflare
The MCP specification introduces a rewritten stateless core alongside updated TypeScript, Python, Go and C# SDKs.
From single call to agents: five new Claude capabilities in Microsoft Foundry (opens in a new tab)
Microsoft
Structured outputs, web search, web fetch, MCP connector and tool search are now available for Claude deployments hosted on Azure.
Research & Papers
Demystifying Agent Skills: Why They Work — Until They Don't (opens in a new tab)
arXiv
A new study looks at when procedural skills actually improve agent behavior and where skill-augmented agents still break across realistic execution environments.
Production & Evaluation
How Uber evaluates AI agents at production scale (opens in a new tab)
Arize / Uber
Uber's experience shows why offline agent evaluation isn't enough: production traces reveal user behaviors, tool disagreements and failure modes that pre-launch datasets don't necessarily capture.
Open Source
Cloudflare OS: an open platform for agents, apps, and work (opens in a new tab)
Cloudflare
Cloudflare open-sourced the agent workspace platform it has been running internally, combining organizational context, skills, isolated runtimes and governed access to internal resources.
Tutorials & Deep Dives
How Google builds, tests and scales Agent Skills (opens in a new tab)
Google Cloud
Google describes its process for managing skills as maintained engineering artifacts, including validation, CI and recurring evaluation rather than treating them as static prompt files.
The runtime behind production Deep Agents (opens in a new tab)
LlamaIndex
ExtractBench evaluates 14 extraction systems across 370 enterprise documents and 4,869 pages, measuring factors including completeness, grounding and cost. Its dataset and evaluation harness are publicly reproducible.
Enjoyed this issue?
Get the next one straight to your inbox.