The Routing Bottleneck: Data, APIs, and Truthful Agents

Research

Random Graphs as Stress Tests for Model Pipelines

Pressure-test your agents with random ugly graphs. This paper loads pipelines with non-IID, noisy data and shows why "it worked on MNIST" is not a valid production strategy. Worth a skim for anyone mapping an agentic system right now.

Routing over RAG: Optimizing the Search Step

Query routing is the invisible bottleneck nobody thinks about. This paper shows that if you treat routing as an optimization problem, you can skip the expensive model calls. A solid way to save budget without sacrificing accuracy on your RAG stack.

Evaluating CoTs: Actions Speak Louder Than Thoughts

Stop trusting your agent's hidden thoughts. This paper argues that "displayed thoughts" are unreliable, which is bad news for tooling that relies on the regex. You need look at the action trace (the actual tool calls) to evaluate agent quality.

News

Anthropic Out-Monetizes OpenAI for the First Time

Inference is becoming a commodity game at the LLM layer, but the data layer is where the money is. If they can combine their chatbot usage data with RAG infra, they could be the enterprise backbone model providers need.

Firecrawl Launches APIs for Agents

This is the professionalization of computer-use. Move from "manipulate pixels" to "hit a structure API." This is how you turn a demo into a product that doesn't break every time a CSS class changes. Build on this before the incumbents own the pipes.