Optimize Agentic Workflows with OpenRouter, HF Benchmarks, and GLM-5.2

The latest updates focus on optimizing agentic workflows and evaluating open-source model performance. OpenRouter introduces specific configuration requirements for the OpenAI Codex CLI, while Hugging Face releases a benchmarking harness designed to measure agent efficiency across different library revisions. In parallel, new open-weight models like GLM-5.2 are challenging existing leaderboards, and OpenAI demonstrates the practical application of autonomous AI in medicinal chemistry research.

1. Codex CLI Configuration for OpenRouter

• OpenRouter now supports the OpenAI Codex CLI through a custom provider block in ~/.codex/config.toml.
• The configuration requires model_provider = "openrouter", a specific base_url, and the critical wire_api = "responses" setting.
• Users must pin a full OpenRouter slug (e.g., openai/gpt-5.3-codex) to avoid model_not_found errors.
• Impact: Developers can route Codex through 300+ models with consolidated tracking, but must update their config to use the Responses API protocol.

2. Hugging Face Agent Benchmarking Harness

• Hugging Face published agent-eval, a tool to measure agent effort (tokens, time, errors) rather than just final accuracy.
• The harness evaluates libraries like transformers across different model sizes and revisions using bare, clone, and skill tiers.
• Testing revealed that adding a CLI and Skill to transformers reduced work for large models but increased token usage for smaller models cloning the repo.
• Impact: Library maintainers can now quantify the impact of API changes on agent efficiency, identifying trade-offs between documentation clarity and model capability.

3. GLM-5.2 Open Weights Release

• Z.ai released GLM-5.2, a 753B parameter text-only model with a 1 million token context window, under an MIT license.
• The model leads the Artificial Analysis Intelligence Index and ranks second on the Code Arena WebDev leaderboard.
• Benchmarks indicate the model is token-heavy, using 43k output tokens per task compared to 26k for GLM-5.1.
• Impact: Developers have access to a leading open-weight text model for long-context tasks, though they must account for higher inference costs.

4. OpenAI Autonomous Chemistry Research

• OpenAI partnered with Molecule.one to use GPT-5.4 and the Maria lab agent to optimize Chan-Lam coupling reactions.
• The system identified TEMPO as an effective additive, improving yields for 88% of boronic acids and 83% of sulfonamides.
• The workflow involved AI-generated hypotheses, automated high-throughput experimentation, and human validation at bench scale.
• Impact: The project demonstrates a viable pattern for AI-assisted experimental design in scientific discovery, highlighting the role of human oversight in validating AI proposals.

Sources

• OpenRouter: https://openrouter.ai/blog/tutorials/codex-cli-openrouter/
• Huggingface: https://huggingface.co/blog/is-it-agentic-enough
• OpenAI: https://openai.com/index/ai-chemist-improves-reaction
• SimonWillison: https://simonwillison.net/2026/Jun/17/glm-52/#atom-everything