Richer AI Gateway traces, full gateway coverage for the Anthropic and Google GenAI SDKs, seven insurance-agent evals, and a wave of Observe polish

Gateway traces now carry your own metadata, request fields, and headers

The AI Gateway now exports caller metadata, selected request-body fields, and allowlisted headers as span attributes, so gateway traces carry more of the context you need to filter and analyze them. Your own dimensions become filterable attributes on the trace instead of an opaque blob. (PR #2230)

What’s new

• Filter by what matters to your business. Tenant, team, feature, or experiment each become their own field on the trace, so you can narrow to exactly the traffic you care about in a couple of clicks.
• Analytics without the workarounds. Group and chart gateway traffic by your own dimensions instead of digging through one long block of metadata.
• Cost and usage by application layer. Every gateway call already records its tokens and cost. Now it also carries the tenant, feature, or team behind it, so you can see what each part of your application spends.
• The full picture of each request. The request settings your application sends travel with the trace too, and you choose which request headers to include. Sensitive headers such as authorization are never exported.

How to turn it on

• Future AGI Cloud. It is on by default. Tag your gateway requests with the metadata you care about, such as tenant or feature, and each one shows up on the trace. See custom properties for how to send them.
• Self-hosted. Turn on OpenTelemetry export in your gateway settings and point it at Future AGI. Metadata and request settings come through by default, and you can pick which request headers to add.

Why it matters

Filtering gateway traffic by your own dimensions, such as tenant, feature, or experiment, used to mean matching substrings inside one JSON string. Now each dimension is a first-class attribute you can filter, group, and chart on.

Who it’s for

Teams routing production traffic through the AI Gateway, on cloud or self-hosted, who want to slice traces and cost by their own business dimensions.

Anthropic and Google GenAI SDK calls get full gateway coverage

Calls made through the Gateway with the Anthropic SDK (/v1/messages) or the Google GenAI SDK now run through the same pipeline as the rest of the Gateway: full traces, cost attribution, and your budget and rate-limit controls. These two endpoints are now in line with the OpenAI-format endpoint. (PR #2231)

One set of controls across every SDK. Whether your application speaks OpenAI, Anthropic, or Google GenAI, the gateway applies the same tracing, cost attribution, budgets, rate limits, and plugins, so nothing slips past your guardrails because of the wire format it used.

Improvements

Evaluation

Seven insurance-agent system evals. Seven built-in system evaluators for insurance agents are now in the catalog, ready to add to eval and simulation runs without writing custom criteria. They cover misselling, advice boundaries, lead qualification, claim intake, intake field accuracy, and voicemail detection and handling on the call audio.

Re-run only the errored evals, or the whole task. From an eval task you can now re-run just the evals that errored, or re-run the entire task, from the task detail.

Eval task detail fixes. Opening an eval task whose saved mapping value is a JSON object now loads cleanly, with the value handled as invalid input instead of taking down the dashboard. The eval-task time window shows the preset you picked and labels custom ranges correctly.

Dataset reasoning toggle keeps its state. In a dataset, the Show/Hide Reasoning option on an eval column’s menu now remembers your choice, so reasoning you hide stays hidden when the column expands.

Dataset column reorder and hide stick. Reordering or hiding dataset columns now takes effect and stays that way.

Ground Truth visibility on Observe evals. When Ground Truth is enabled on a template but cannot be applied to a given Observe eval run, that is now surfaced to you rather than the run proceeding silently without calibration.

Observe

Every Error Feed cluster shows its source. Every cluster card in the Error Feed now shows where it came from, the built-in error taxonomy or a specific eval task you configured, so it is clear at a glance which findings are surfaced by Future AGI and which come from your own evals.

Custom attribute filters: every key, faster matches. The attribute discovery behind filter dropdowns now reads keys exhaustively rather than sampling rows, so projects with many distinct span attributes see the full list to filter on, and custom attributes stay in the picker. Exact-value filters on custom attributes in Observe and on dashboard graphs now resolve faster and more consistently, including on unicode values.

Total Tokens on the Sessions grid. Total Tokens is available as a filter and column on the Sessions grid, and the column toggle now persists.

Alerts: faithful editing, working email links, clearer charts. Reopening a saved alert now shows exactly what was stored, including the metric, threshold, and frequency, so saving from the editor keeps your real configuration. The Open monitor button in alert emails now opens the alert that fired through a region-correct link, and alert config charts default to a 7-day window with readable tooltips in dark mode.

Clearer sampling rate control. The Error Feed sampling-rate control now states what it controls, so it is distinct from a similarly labeled eval setting.

Observe graph metric picker scoped to the project. The metric picker on the Observe graph, across the Traces, Spans, Sessions, Users, and Compare views, now lists only the evals and annotation labels in the current project rather than every one in the organization.

Voice Observe: quick filters and distinct saved views. Quick filters are now available in voice Observe, so you can narrow to the calls you care about without building a full filter expression. Voice observability now rejects a saved view whose name is already in use, so your saved views stay distinct.

Dashboards

Dashboard polish. Dashboard pies now render one chart per metric and only show a pie when a breakdown is set, so the chart matches the data. Chart line colors are now consistent for a series across a widget, and a widget’s description is shown on the individual widget.

Agents

Falcon shows a connector’s discovered tools. The Falcon Customize panel now lists the tools discovered from a connected connector, so you can see exactly what an agent can call.

Platform

Open-source build fixes. A large batch of fixes for the open-source build: gateway API-key field styling, the agent-scenario dark-theme popup, action-button styling, default-tag visibility, Hugging Face import, datapoint row selection, and more.

AI Gateway

Anthropic server tools on the OpenAI-format endpoint. Anthropic server-side tools, such as web search, now work when called through the Gateway’s OpenAI-format endpoint, so Claude searches when you ask it to.