Splunk aims to show whether AI agents are earning their keep

Résumé détaillé

New Tokenomics capabilities will connect agent costs with output quality and business measures, while smaller evaluation models aim to make continuous checks more affordable.

New Tokenomics capabilities will connect agent costs with output quality and business measures, while smaller evaluation models aim to make continuous checks more affordable.

An AI agent can generate plenty of activity without delivering much value. Splunk’s latest observability announcements aim to help organisations see the difference by bringing the cost of running agents together with assessments of their work and the business outcomes they are meant to support.

Unveiled in conjunction with Cisco Live in Melbourne, the developments expand Splunk Agent Observability with enhanced Tokenomics for AWS environments, new Eval-G1 evaluation models, and an Evaluation Tuner that lets organisations adapt those assessments to their own requirements.

The evaluation models and tuner will become generally available during November 2026. The enhanced Tokenomics capability is scheduled for private preview in December 2026 and general availability through AWS Marketplace in early 2027.

Tokenomics builds on cost attribution introduced at Splunk’s .conf26 event, which tracks token spending across AI agents and coding tools including Claude Code, Codex, and Cursor.

The new capability will combine usage information from Amazon Bedrock and Amazon CloudWatch with evaluations of an agent’s work and customer-defined business indicators. For example, Splunk says Net Promoter Score (NPS) and cost per claim are measures that can be brought into the reporting.

This all helps when examining AI expenditure. A relatively expensive agent might justify its cost by completing valuable work reliably, while a cheaper agent might create extra expense if its outputs need frequent correction or its failed attempts repeatedly consume resources.

It's these distinctions that Splunk wants to help expose. It's one thing to count tokens, but another to understand value.

The need is reflected in Splunk’s State of Observability 2026 research. In the global survey, 55% of respondents said they could not control AI…

Début de l'article, extrait de iTWire. Pour lire la suite, ouvrez l'article chez le média.