AI cost routing, agent failure diagnosis, and a CNC manufacturability checker

If you're running AI agents in production — chatbots that call APIs, document processors, internal tools — you're probably defaulting every call to a big, expensive frontier model. Switchcraft is a new model router built specifically for agentic tool-calling. Instead of sending every request to the most capable (and priciest) model, it evaluates each call and routes it to the cheapest model that can handle it reliably.

Existing routers were designed for chat completions, not tool use. Switchcraft is the first router optimized for the structured, function-calling patterns that agents actually use. The result: lower per-call costs without meaningful accuracy loss on the tasks that matter.

For a mid-market company running AI workflows at scale — say, processing hundreds of supplier invoices or routing customer service tickets — this kind of routing is a direct line item on your AI budget. It's the difference between paying premium rates on every call and paying premium only when you need to.

When AI agents go wrong in enterprise workflows, diagnosing the failure is brutal. Did the agent skip a required tool call? Fire an unnecessary one? Take an action whose consequences didn't show up until three steps later? Current observability methods are mostly external — they watch inputs and outputs but don't show you what happened inside the decision.

A new paper proposes interpretability tools built specifically for agent tool-use failures. The framework gives you visibility into why an agent made (or didn't make) a specific tool call, rather than just showing you that something went wrong after the fact.

This matters for any business deploying AI in processes where mistakes are expensive — think order management, compliance checks, or financial reconciliation. If you can't diagnose failures, you can't trust the system. And if you can't trust the system, you're paying for a human to watch the AI, which defeats the point.

SREGym is a new benchmark that drops AI agents into realistic, live infrastructure failure scenarios — the kind of outages and incidents that cost real businesses real money. This isn't a toy problem set. It simulates the messy, high-pressure conditions of actual site reliability engineering (SRE — the discipline of keeping production systems running).

Early results show frontier models can handle some incident types but still struggle with the multi-step diagnostic reasoning that experienced human engineers do instinctively. The benchmark is designed to be extensible, so it should track agent improvement over time.

If you've been wondering whether AI can take on real ops work — not just answer questions about it — this benchmark is how the industry will measure progress.

MachinaCheck is a multi-agent system that evaluates whether a part design can actually be CNC machined. It analyzes geometry, material constraints, and tooling requirements — the kind of checks that usually require an experienced machinist or manufacturing engineer to eyeball.

What makes this notable beyond the use case: it runs on AMD MI300X hardware, not the NVIDIA stack that dominates AI infrastructure. That's a signal for companies watching GPU costs — competition on the hardware side eventually means lower prices for everyone running AI workloads.

For manufacturers specifically, this is a concrete, non-hype example of AI doing useful shop-floor work. Not generating marketing copy. Not summarizing emails. Checking whether a part can be made.

AIDA (Autonomous Insight Discovery Agent) is a new framework designed for something every mid-market company with multiple systems hits immediately when they try AI-powered analytics: getting useful analysis out of complex, siloed enterprise databases.

The paper addresses the specific pain points — messy schemas, dynamic SQL generation that breaks, and the need for multi-dimensional analysis across tables that were never designed to talk to each other. AIDA is an end-to-end system that goes from raw enterprise data to structured business insights without requiring a data engineer to hand-hold every query.

If you've tried plugging an LLM into your ERP or accounting system and gotten garbage back, this research is aimed squarely at that problem.