Open Source20d ago
OpenUI 1.0 ships: the open-source generative-UI standard adds custom actions and functions; its own benchmark shows ~52% fewer tokens than JSON render formats
On 2026-10-15 Thesys shipped OpenUI 1.0 (MIT, ~11K GitHub stars). OpenUI Lang 1.0 adds custom actions (defineAction), custom functions (defineFunction), library.extend and a stored-message format, so agents can stream interactive cards, forms and charts instead of walls of text. In the project's tiktoken benchmark, seven UI scenarios total 4,800 tokens, 52.8% fewer than Vercel JSON-Render. The 1.0 features currently require the React renderer.
Agents29d ago
Judged Useless, Queried Anyway: NTU Reveals Fatal Decoupling between Evidence Evaluation and Stopping in Tool-Using Agents
When an external tool or search endpoint repeatedly returns irrelevant noise, a rational agent should cease querying and act upon available knowledge. A comprehensive study by Nanyang Technological University (NTU) and A*STAR (arXiv:2610.06191) uncovers a profound behavioral failure mode: while LLM agents correctly identify tool outputs as useless 97% to 100% of the time, they systematically fail to translate those negative evaluations into stopping decisions, entering persistent query loops. Analyzing seven frontier open-source and proprietary models under controlled tool failure modes, the authors show that prompt warnings regarding token budgets, call costs, or explicit stopping guidelines fail to enforce rational early exits—often prompting smaller 7B-8B models to burn calls until hit by hard deadlines. Evidence-grounded stopping emerges reliably only when the execution harness enforces a programmatic policy: mandating final generation after five consecutive self-judged useless queries. This integration rule raises task completion rates across every evaluated model during source failures, confirmed via a 300-question pre-registered replication trial.
Tooling39d ago
CopilotKit 1.74.0: consume learned skills across SDK containers; Web Inspector targeted What's New + HUD telemetry
CopilotKit shipped v1.74.0 on 2026-09-25 (npm @copilotkit/react-core 1.74.0): SDKs can consume learned skills from multiple containers, expanding skill assets beyond a single runtime; Web Inspector gains targeted What's New notifications and launcher HUD impression/action tracking. Fixes: threads drawer no longer locks on unresolved entitlements, stop refreshing realtime credentials for sockets that never open, stop per-message state cloning / virtual-scroll tug-of-war in React, and a Memory policy that grants nothing no longer fails the run.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.