It seems I can't look at the internet without seeing talk of Ralph Wiggum or Gas Town. Developers are either running their coding agent in loops or strapping together many parallel coding agents. Coding agents are far and away the most successful agents out there right now, so I always think it's worth keeping an eye on. It's fun to see multiple sub-agents, roles, memory and coordination on one side and a loop on the other.

This week in AI++ we have articles on building agent memory, evaluating your agents, and some different agent design patterns. Also, see how ServiceNow shipped a very vulnerable agent (it wasn't the AI's fault), and find out how Nano Banana got its name.

​Phil Nash​
Developer relations engineer for Langflow​

🛠️ Building with AI, Agents & MCP

Agent architectures and patterns

The LangChain blog published a great round-up of multi-agent architectures, including the trade-offs you make between them. For each of those agents, check out the design patterns you can use to get them to work well, from giving them a computer to different ways to handle context. Vercel have been arguing that complex agents aren't required, they can be great as long as they have access to a filesystem and bash. Although if you're working with smaller models there's some work to do to make them smarter for your use-case.

Memory is hot

OpenAI published a very in-depth tutorial on building long-term memory using the Agents SDK. Meanwhile GitHub described how they have added memory to Copilot. Copilot specifically includes links to lines of code as citations for its memories, meaning it can correct or update itself if things change.

Building your own coding agents

The Claude Agent SDK is the engine behind Claude Code, and Nader Dabit uses it to build a code review agent. On the Langflow blog, we look into a small part of the coding lifecycle, generating commit messages using a git MCP server.

🗞️ Other news

• Google, Shopify, Etsy, and others have been working on the Universal Commerce Protocol to give agents the ability to buy things for you. Here's how UCP works under the hood​
• OpenAI released the Open Responses spec, a shared schema based on their Response API and supported by NVIDIA, Ollama, LM Studio, and more
• At Block they've been working on red-teaming Goose, their open source agent. It's a great look into adversarial attacks on agents
• ServiceNow released an agent that was all kinds of vulnerable. The kicker? The vulnerabilities were around authentication, identity and excess privileges in the functions that the agent was given access to, not in the agent itself. Check out Snyk's breakdown of the issues​
• Google is working on adding gRPC as a transport for MCP​
• Ever wondered how Gemini's Nano Banana got its name? Wonder no more​

🧑‍💻 Code & Libraries

• ​Tool UI and JSON Render are two projects that provide schemas for UI components so that LLMs can produce output that will generate UIs
• ​Headroom provides various tools to compress tool content and context to give LLMs more space
• Vercel's love for CLIs for agents produced Agent Browser, a CLI that can power a headless browser
• On the same note, MCP-CLI is a CLI that allows for dynamic discovery of MCP

📅 Events

On January 29th IBM is hosting AI Demystified, a virtual event and hackathon covering many different open-source AI tools including Langflow, CUGA, Granite models, and the Agent Lifecycle Toolkit (ALTK).