In the news
Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation
arXiv cs.AI · Published · 3 min read
In 30 seconds
- What happened
- Research shows that local serving stacks like Ollama and vLLM significantly affect measured tool-use performance in coding agents, not just model behavior.
- Why it matters
- Matters for engineers evaluating local AI models on tool-calling tasks, especially when comparing results across different serving frameworks.
- Watch out
- Different serving stacks handle the same request differently; rejection and retry failures may be misclassified as model failures, inflating error rates.
- agent
- tool-use
- inference
- serving
- eval
The patterns behind this
- Context Editing & Tool-Result Clearing
- Constitutional AI Evaluation Framework
- AISI Evaluation Framework
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.