Extend Operate > Assets cost estimation to external/registered agents #448
Unanswered
Aditya Bhaskar (InspectorAB)
asked this question in
Product Feedback and Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Technical Feedback
The Estimated cost tile under Operate > Assets > Agents documents that its estimates don't include prompt agent and non-Foundry agent costs. So native Foundry agents get token usage, budget consumption, and cost visibility in the Assets pane by default, but agents registered externally via OTEL (same agent ID as the code agent, calling their own LLM outside Foundry Models) don't get this at all.
Desired Outcome
Let an external agent registered under Operate > Assets declare which LLM/model it's calling, so Foundry can dynamically estimate token consumption and cost the same way it does for native agents, and surface it in the same Assets view alongside Foundry-native agents. This doesn't need real-time metering precision even a self-reported model + usage estimate, tracked in the same place over time, would close most of the visibility gap and give teams one place to compare cost across both native and external agents long-term, instead of stitching together separate APIM/App Insights pipelines just to answer "what is this agent costing us."
To be clear, declaring the model is metadata for cost estimation only it doesn't imply the call routes through Foundry Models. The agent may still be calling OpenAI, Bedrock, a self-hosted model, or running in another cloud entirely, for reasons like framework choice, data residency, or provider pricing outside Foundry's catalog. It just reports its model and usage back to Foundry for observability, the same way it already reports traces via OTEL.
Current Workaround
Teams are building their own metered gateways to get this: routing external agents through APIM, extracting token counts, estimating cost against reference model pricing, and logging everything as custom dimensions in Application Insights, then writing KQL queries to aggregate cost per agent per model. This works, but it means every team re-implements the same cost-tracking pipeline by hand outside Foundry, instead of getting it as part of the same Assets view Foundry-native agents already have.
All reactions