Manu Shukla

Posted on Aug 4 • Originally published at ecorpit.com

2 new Gemini Enterprise trace spans that pinpoint a slow agent call (2026)

#geminienterprise #observability #opentelemetry #cloudtrace

2 new Gemini Enterprise trace spans that pinpoint a slow agent call (2026)

Summary. On 4 August 2026 Google extended tracing across the whole Gemini Enterprise data-connector workflow and added two spans: execute_tool, which covers execution of a tool on the agent orchestration layer, and invoke_connector, which covers request logic and execution on the connector execution layer. Together they close the gap between "the assistant felt slow" and "the Jira Cloud call took 4 of those seconds". Spans land in Cloud Trace inside your own Google Cloud project, where the default retention is 30 days and ingestion is metered at $0.20 per million spans after the first 2.5 million per billing account. The same release note also confirmed that Gemini 3.5 Flash was removed from the global region in the Gemini Enterprise app on 4 August 2026, so a latency regression dated that week may be a model change rather than a connector fault. Context for why this matters: PwC's AI Agent Survey found 79% of organisations have adopted AI agents, and Braintrust raised an $80 million Series B on 17 February 2026 at an $800 million valuation on the argument that a single agent interaction can now emit hundreds of megabytes of trace data. Three span types already existed in the agent Traces tab — Agent to Tool, Invoke Agent and Agent to Model. The two new ones make the fourth hop, the one leaving Google's network, visible.

This is a configuration and debugging guide, not a product tour. It assumes you run Gemini Enterprise Standard, Plus or Frontline, that you have at least one third-party connector wired up, and that someone has asked you why a prompt takes eight seconds.

What actually changed on 4 August 2026

Before this release, a Gemini Enterprise trace stopped being useful at the agent boundary. You could see that an agent invoked a tool. You could not see what happened inside the connector that carried the request to Zendesk, ServiceNow or Confluence Cloud, so a slow answer looked identical whether the model was thinking, the orchestration layer was queuing, or a third-party API was rate-limiting you.

The Gemini Enterprise release notes describe the change plainly: tracing support is extended end to end across the data connector workflow, with two new trace spans. execute_tool represents the execution of a tool on the agent orchestration layer. invoke_connector represents the request logic and execution on the connector execution layer. Google's stated purpose is to let you visualise the end-to-end parent-child relationship from the assistant prompt through to the third-party API.

Three query paths come with them. You can search and filter for the spans in Trace Explorer by service or by span name. You can query them by turn-level legacy assist tokens through the gemini_enterprise.assist_token attribute. You can look a trace up by its W3C trace ID, which is what makes it possible to correlate a Gemini Enterprise trace with a trace your own backend already emits.

That third path is the one platform teams should care about most. A W3C trace ID is the join key between Google's view of the agent and your view of the downstream service. Without it you are eyeballing two timelines and hoping.

Turn instrumentation on before you look for the spans

Tracing is not on by default, and the toggle is not in one place. Per Google's access traces and spans documentation, you need three things before a single span appears.

First, roles. You need the Gemini Enterprise Admin role or the Google Cloud console Gemini Enterprise User role, plus the Cloud Trace User role (roles/cloudtrace.user) to read traces at all. A common failure at this step looks like an empty Trace Explorer rather than an error, which wastes an afternoon.

Second, an existing Gemini Enterprise web app.

Third, the observability setting Enable instrumentation of OpenTelemetry traces and logs. Where you set it depends on the agent type. For the Core Assistant agent you enable it in app-level configuration. For Agent Designer employee-made agents and Deep Research agents you enable it inside the agent's own Observability tab. Teams that flip the app-level switch and then wonder why a custom agent emits nothing have usually missed the per-agent toggle.

There is a second switch next to it, and it deserves a decision rather than a default. Google's manage observability settings page describes Enable logging of prompt inputs and response outputs as logging the full content of user prompts and responses into Cloud Logging, including sensitive data or personally identifiable information. Google's own caution on the page is direct: restrict log access to authorised personnel only. Turning the OpenTelemetry setting off also turns this one off and stops collection of traces, spans, span logs and metrics.

The engineering judgement here is simple. Turn on instrumentation everywhere. Turn on prompt and response logging only in a non-production app, or in a production app for a fixed, ticketed debugging window with an owner and an end date.

Reading the waterfall: what each span tells you

The agent Traces tab is available for the Core Assistant agent, Agent Designer employee-made agents and Deep Research agents. Its summary table lists status, trace span ID, span name, span type, duration, input, output and start time. Clicking a span ID opens a detail page with a Summary block, a toggle between Graph and Timeline views, a View in Cloud Trace link, and per-span sections for Logs, Evaluation, Input & Output, Logs & Events, Stacktraces, Metadata & Links, and Attributes.

Two quirks are worth knowing before you read your first trace. Status currently displays as Unset for every span, so it is not a health signal — a failing call shows up in the Stacktraces section, not in the status column. And the summary table reports duration in minutes while the selected-span detail reports it in seconds, which makes a fast trace look like a rounding error until you drill in.

Span or span type	Layer it covers	What a long one usually means	Where to look next
`Invoke Agent`	Agent invocation	Orchestration overhead, or a sub-agent doing real work	Child spans in the Graph view
`Agent to Model`	Agent to a large language model	Model latency, long context, or a model swap	Model configuration; check release notes for regional model changes
`Agent to Tool`	Agent to an external tool	The agent chose a tool and waited	The new `execute_tool` child
`execute_tool`	Agent orchestration layer	Tool dispatch, argument handling, retries inside orchestration	Attributes and Logs & Events on the span
`invoke_connector`	Connector execution layer	Request construction plus the round trip to the third-party API	Third-party API status page and rate limits
Span logs	Timestamped events inside a span	Retry storms, auth refresh loops, pagination	Logs & Events, then Cloud Logging

The diagnostic value comes from the difference between two spans rather than either one alone. If execute_tool runs long but invoke_connector is short, the third-party API is fine and your orchestration is the problem. If invoke_connector dominates, stop tuning prompts and go read the vendor's rate-limit documentation.

Three latency shapes you will actually see

After the parent-child structure is visible, most slow Gemini Enterprise turns fall into one of three shapes.

The first is model-dominated. Agent to Model owns most of the trace and the connector spans are noise. This is the shape to expect the week of 4 August 2026 for anyone who was pinned to Gemini 3.5 Flash in the global region, because the release notes record its removal from that region on that date. Check the model before you blame the plumbing.

The second is connector-dominated. invoke_connector owns the trace. Either the third-party system is slow, or your connector is paginating through more data than the answer needs. Narrowing the data store or adding filters is usually cheaper than any prompt change.

The third is fan-out. No single span is slow, but there are many of them. This is the shape that used to be invisible, and it is the reason the two new spans matter: a dozen sequential execute_tool and invoke_connector pairs at 400 milliseconds each is a five-second answer that no individual span explains.

What tracing costs

Trace data is stored in Cloud Trace in your own Google Cloud project, and retention is governed by Cloud Trace — 30 days by default, after which data older than the retention period is deleted automatically. The bill follows Google Cloud Observability pricing.

Meter	Rate (as of August 2026)	Free allotment
Cloud Trace ingestion	$0.20 per million spans	First 2.5 million spans per billing account
Cloud Trace scopes	No charge to configure	Not applicable
Cloud Logging ingestion	$0.50 per GiB, includes up to 30 days of storage	Per Google Cloud Observability pricing
Network telemetry log ingestion	$0.25 per GiB	Per Google Cloud Observability pricing
Cloud Logging retention beyond 30 days	$0.01 per GiB per month	Not applicable

Two details change the arithmetic. Cloud Trace charges are based on spans ingested and scanned, and Trace pricing does not apply to spans auto-generated by App Engine Standard, Cloud Run functions or Cloud Run — those ingestions are non-chargeable. So a Gemini Enterprise agent hosted on Cloud Run contributes chargeable spans from Gemini Enterprise itself while its Cloud Run hosting spans ride free.

The span meter is rarely what hurts. At $0.20 per million spans, an agent emitting eight spans per turn would need roughly 312,500 turns to consume the 2.5 million free allotment, and a further million spans costs less than a single seat of most observability tools. The line that grows is Cloud Logging, because prompt and response logging writes full payloads at $0.50 per GiB. That is the real reason to keep the second toggle off outside a debugging window.

Native tracing versus a dedicated agent observability platform

Gemini Enterprise tracing answers "where did the time go". It does not answer "was the answer any good". If you need scoring, regression gates in CI, or a shared workspace for non-engineers, you are buying a second tool. Braintrust's June 2026 comparison of agent observability platforms lists the current entry prices.

Option	Entry price (as of June 2026)	What it adds over Cloud Trace	Trade-off
Gemini Enterprise + Cloud Trace	$0.20 per million spans after 2.5 million free	Native connector spans, W3C trace ID correlation, no extra vendor	No output scoring; 30-day default retention
Braintrust	Free tier, Pro $249/month	Evaluation built into observability, GitHub Actions gating	Setup and configuration effort
Agenta	Free Hobby tier, Pro $49/month	Open source, self-hostable, trace annotation	Hobby tier caps at 2 seats and 5,000 traces/month with 30-day retention
Helicone	Free 10,000 requests/month, $20 per seat/month	Proxy-based logging with no code change	Shallower tracing depth than SDK-based tools
Galileo	Free 5,000 traces/month, Pro from $100/month	Low-latency evaluators on live traffic	Smaller community than established platforms
Fiddler	Custom enterprise pricing	ML and LLM governance for regulated industries	Setup suits organisations with dedicated ML teams

The case for adding one is scale, not sophistication. Ankur Goyal, co-founder and Chief Executive of Braintrust Data Inc., put the problem this way when the company announced its $80 million round on 17 February 2026: "Teams have never had less conviction about what will fail next." Multistep agents with tool calls and intermediate reasoning generate enormous volumes of data, amounting to hundreds of megabytes per interaction. Cloud Trace will happily store that. It will not tell you the answer got worse.

Our own read: start with the native spans. Most teams have not yet exhausted what a correctly instrumented Gemini Enterprise trace tells them, and a second observability bill before the first one is understood buys dashboards, not answers.

India-specific considerations

Two things matter for teams running Gemini Enterprise from India.

The first is where the traces and the model run. The release notes record that from 30 June 2026 the Gemini Enterprise app can be used in the India (IN) and Singapore (SG) regions with at-rest data residency and machine learning processing in region, with the latest Gemini 3.5 Flash model usable in those regions under the same in-region terms. Japan (asia-northeast1) and the United Kingdom (europe-west2) followed in July 2026. These locations are generally available with an allowlist, so access goes through your Google account team, and certain limitations apply. The 4 August 2026 removal of Gemini 3.5 Flash applied to the global region, which is a different configuration from an in-region deployment — check which one your app is on before you assume you are affected.

The second is what you write to logs. Enabling prompt and response logging puts the full content of user prompts and responses into Cloud Logging, and Google's documentation states plainly that this includes sensitive data or personally identifiable information. Under India's Digital Personal Data Protection Act 2023, that is personal data being processed and retained, and it needs the same purpose limitation, access control and deletion discipline as any other store. Cloud Logging's default is up to 30 days of storage included with ingestion, with $0.01 per GiB per month beyond that — a retention decision, not an accident. Teams working through this alongside broader obligations usually treat it as part of the same programme as their enterprise AI agent governance layers.

A practical rollout order

Do these in sequence rather than in parallel; each step tells you whether the next one is worth doing.

Grant roles/cloudtrace.user to the two or three people who will actually read traces, plus the Gemini Enterprise Admin role where needed.
Enable Enable instrumentation of OpenTelemetry traces and logs at app level for the Core Assistant, then separately in the Observability tab of each Agent Designer and Deep Research agent.
Leave prompt and response logging off. Open a ticket for the first time someone genuinely needs it.
Run ten representative prompts and open the Traces tab. Confirm you see execute_tool and invoke_connector under the tool calls. If you do not, you missed a per-agent toggle.
Capture the W3C trace ID for one slow turn and search for the same ID in your own backend tracing. That correlation is the whole point of the release.
Only after the above, decide whether output quality needs an evaluation platform on top. Cost it against the tables above rather than against a demo.

Teams that have already done this for other stacks will recognise the pattern from OpenTelemetry GenAI conventions for LLM agent tracing, and the same instrumentation discipline shows up in every serious enterprise AI agent production deployment. If you are budgeting the whole platform rather than the traces alone, the Gemini Enterprise governance and billing controls guide covers the seat and token side.

FAQ

What are the two new Gemini Enterprise trace spans?

Google added execute_tool and invoke_connector on 4 August 2026. The first represents execution of a tool on the agent orchestration layer. The second represents request logic and execution on the connector execution layer. Together they make the path from an assistant prompt to a third-party API visible as parent-child spans.

Do I need to enable anything before traces appear?

Yes. You need the Gemini Enterprise Admin or Gemini Enterprise User role, the Cloud Trace User role, an existing Gemini Enterprise web app, and the observability setting called Enable instrumentation of OpenTelemetry traces and logs. That setting is app-level for the Core Assistant and per-agent for Agent Designer and Deep Research agents.

How long are Gemini Enterprise traces kept?

Traces and spans are stored in Cloud Trace inside your own Google Cloud project, and retention follows Cloud Trace. The default retention is 30 days, and data older than the retention period is deleted automatically. Google directs readers to the Cloud Trace quotas and limits documentation for the authoritative current values.

What does Gemini Enterprise tracing cost?

Cloud Trace ingestion is $0.20 per million spans, with the first 2.5 million spans per billing account free, and configuring trace scopes carries no charge. Cloud Logging ingestion is $0.50 per GiB including up to 30 days of storage, then $0.01 per GiB per month for longer retention.

Should I turn on prompt and response logging?

Only for a bounded debugging window. Google's documentation states that this setting logs the full content of user prompts and responses into Cloud Logging, including sensitive data or personally identifiable information, and cautions that log access should be restricted to authorised personnel. Under India's DPDP Act 2023, that content is personal data.

Why did my Gemini Enterprise latency change in early August 2026?

Check the model before the connectors. The release notes record that Gemini 3.5 Flash was removed as a model from the global region in the Gemini Enterprise app on 4 August 2026. An app running in an in-region deployment such as India or Singapore is a different configuration from the global region.

Can I correlate a Gemini Enterprise trace with my own backend traces?

Yes, through the W3C trace ID. Google documents three ways to find the new spans: searching and filtering in Trace Explorer by service or span name, querying by turn-level legacy assist tokens through the gemini_enterprise.assist_token attribute, and looking traces up by W3C trace ID. The last is the join key for cross-system correlation.

Does native tracing replace an agent observability platform?

No. Cloud Trace shows where time was spent; it does not score whether the answer was correct. Platforms compared by Braintrust in June 2026 start at a free tier, with Agenta Pro at $49 per month, Helicone at $20 per seat per month, Galileo from $100 per month and Braintrust Pro at $249 per month.

How eCorpIT can help

eCorpIT is a senior-led engineering organisation in Gurugram, CMMI Level 5, MSME Certified and ISO 27001:2022 certified, and we build and instrument production agent platforms on Google Cloud. We wire up Gemini Enterprise observability end to end: roles, per-agent instrumentation, W3C trace ID correlation with your existing backend tracing, and a logging policy that keeps prompt payloads out of long-lived buckets. We design these applications aligned with DPDP Act 2023 requirements, and we cost the trace and log meters before you turn anything on. Tell us where your agent is slow at /contact-us/ and we will start from your traces.