Salesforce added continuous agent evaluation and hard spending caps to Agent Fabric, which it says now handles 1.2 billion requests a day
AspireAI community · Salesforce
Salesforce announced today that Agent Evals, API Evals and advanced cost controls have arrived in Agent Fabric, its control plane for managing agents across platforms. Agent Evals generates tests before deployment and then evaluates agents, LLMs and MCP servers in production against defined quality thresholds, triggering predefined interventions - including an Agent Kill Switch - when results fall short. API Evals generates functional and performance tests from API specifications to catch breaking changes before they reach production. On the cost side, a FinOps Dashboard and Model Wallet give budget owners hard limits, and Cost Optimization Policies cut token usage by up to 75 percent based on test scenarios through semantic caching, payload compression and response trimming.
Salesforce cites fewer than half of organisations with agents evaluating them on live traffic, and 96 percent of enterprises running agents on more than one platform, as the gap the product addresses. Agent Evals and API Evals enter beta in November 2026 with general availability planned for January 2027; the Custom Agent Scanner and Triggered-Agent Kill Switch reach general availability in January 2027, while the Kill Switch, Open Policies for Federated API Gateways, FinOps Dashboard, Model Wallet and Cost Optimization Policies are generally available as of today.
The honest limit: the 75 percent token reduction is Salesforce's own figure from test scenarios rather than from customer production telemetry, and the headline evaluation features are still two months from general availability. Salesforce's own availability note says the article may reference services still in development and unreleased.
https://www.salesforce.com/news/stories/agent-fabric-catch-failures-control-ai-costs
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.