New tools for agents and testing their weaknesses

This roundup includes a local assistant for researchers and ways to prepare data for enterprise agents. Studies examine when specialized prompts help and where models can do without them.

➡️ News:

🔹 K-Dense BYOK released a local AI assistant for researchers: This is an open-source project that runs on the user’s computer and allows them to choose their own model. 🔹 ServiceNow AI introduced AutoSynthData for generating data for enterprise agents 🔹 Praxa separates an agent’s proposal from a confirmed action: In the project’s description, execution proceeds through authorization checks, external-result verification, and approval.

➡️ Useful materials:

🔹 Study: specialized system prompts for scientific tasks increase costs without a consistent improvement in accuracy: The authors reviewed 503 profiles for different scientific professions. 🔹 Study examines whether small models can handle routine tasks in an agentic system: For example, tool selection, memory recording, or preliminary command approval. 🔹 Study finds a weakness in humor evaluation for training language models: The surprise metric considered shuffled words to be successful jokes. 🔹 Study explores when causal world models help modular agents plan: The authors compare them with models based on observable action sequences.

📝 If you would like to add other news and materials to the list, write in the comments.