Gemini Enterprise Agent Platform Updates & Release Notes

Change

Use a single regional Agent Registry instance for Gemini Enterprise apps

For Gemini Enterprise apps, we recommend associating each Agent Gateway with a single regional Agent Registry instance in the same project and region as the gateway. Use the us-central1 registry for global and us apps, and the europe-west1 registry for eu apps.

This note updates the September 9, 2026 entry. For more information, see Register your agents and destination resources.

Fixed

CodeMender updates (v0.17.0)

This release introduces updates to CodeMender:

• Support bundle diagnostics: Introduce cm support-bundle to collect environment metadata, configuration policies, health reports, and session logs into a compressed archive for streamlined troubleshooting.
• Plain output mode: Add the persistent --plain flag (configurable via the CM_PLAIN environment variable or output.plain in config.yaml) to suppress decorative emojis, ANSI escape sequences, and Unicode borders for compatibility with screen readers and CI/CD log pipelines.
• Dedicated exit codes for setup failures: Assign distinct exit codes (such as code 10 for credential issues) to configuration and authentication failures, preventing CI/CD pipelines from confusing environment setup issues with security findings.
• Tailored CI/CD credential guidance: Provide environment-specific authentication guidance when running inside automated CI/CD systems such as GitHub Actions, GitLab CI, Buildkite, and Jenkins instead of interactive login prompts.
• Single-line JSONL statistics: Add the --jsonl flag to cm stats to output single-line JSON records with model and team attribution for seamless integration with Cloud Logging and log aggregation pipelines.
• Granular sandbox and policy flags: Replace the overloaded --unrestricted flag with orthogonal flags (--sandbox=false) and configurable command policies to decouple container sandboxing from execution safety controls.
• Sandbox configuration protections: Mount .git/hooks, .git/config, and .codemender read-only within writable execution sandboxes during cm fix and cm verify to prevent modifications to repository security policies.
• Bug fixes:
  • Added a security warning when commands are executed with both sandbox isolation disabled and interactive approval skipped.
  • Enforced allowedRoots path boundaries across host discovery and remediation file operations to prevent traversal via dangling symlinks.
  • Supplied exact finding line ranges to remediation agents during cm fix to eliminate manual file scanning overhead.
  • Updated cm report import to properly treat findings without location data as file-level findings rather than defaulting to the first line.
  • Ensured cm session cancel routes requests to the session's original region regardless of local environment overrides.
  • Fixed finding status tracking in cm verify so unexploitable verdicts update the local state database rather than leaving findings open.

For more information, see the CodeMender documentation.

Feature

Gemini Robotics ER 2 (Preview)

Gemini Robotics ER 2 (gemini-robotics-er-2-preview) is available in Preview. Gemini Robotics ER 2 is an embodied reasoning vision-language model built on Gemini 3.5 Flash for robotics applications, including spatial reasoning, agentic code execution, task orchestration, and video understanding.

Feature

xAI's Grok 4.7 is generally available

Grok 4.7 is now generally available (GA) for production use on the global endpoint and the US multi-region endpoint.

Control user intent matching for semantic governance policies

You can now turn user intent matching on or off for semantic governance policies by setting the sgpUserIntentMatching key on the authorization extension metadata. User intent matching is on by default; set the key to DISABLED to turn it off. Turning it off reduces protection against indirect prompt injection. When the key is set, policy evaluation logs report the mode each evaluation used in the sgpUserIntentMatchingMode field.

For more information, see Control user intent matching.

Cloud Storage output support for Gemini image generation models

Direct Cloud Storage output persistence is generally available (GA) for Gemini Nano Banana 2.1 (gemini-nano-banana-2.1), Gemini 3.1 Flash Image (gemini-3.1-flash-image), Gemini 3.1 Flash-Lite Image (gemini-3.1-flash-lite-image), and Gemini 3 Pro Image (gemini-3-pro-image). You can set outputConfig.outputUriPrefix and generationConfig.responseFormat[].image.delivery to "URI" in generateContent and streamGenerateContent requests to write generated and edited images directly to a Cloud Storage bucket and receive the object URI in fileData.fileUri instead of inline bytes.

For more information, see Save generated images to Cloud Storage.

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are Generally Available (GA)

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are available in General Availability (GA) in the global location and the us and eu multi-regions. The models support structured speaker and style metadata, inline vocal tags, multi-speaker dialogue, and streaming. Voice design and voice replication capabilities remain available in Preview.

For more information, see Generate speech with Gemini TTS.

Semantic governance policies support more model providers

Semantic governance policies now evaluate tool calls from agents that use models from providers other than Gemini. The policy engine reads each provider's request and response format, so the same policies apply whichever model your agent uses. Supported APIs include the Gemini Developer API and Gemini models on Gemini Enterprise Agent Platform, the Anthropic Messages API, Claude models on Gemini Enterprise Agent Platform, the OpenAI Chat Completions API, the Mistral AI API, the OpenAI-compatible endpoint on Gemini Enterprise Agent Platform (including Llama models), and self-hosted OpenAI-compatible servers such as vLLM. Streamed model responses aren't supported.

To govern these providers, update the when condition of your authorization policy to match each provider's request path. The updated condition also matches requests that include a query string.

For more information, see Supported model providers and Connect the policy engine to Agent Gateway.

Fixed

CodeMender updates (v0.16.0)

This release introduces updates to CodeMender:

• Deep scan by default (cm find): cm find now runs a concurrent file-by-file deep scan (--deep) by default for higher vulnerability recall across repositories. To run a single-session quick scan instead, pass the --quick flag (cm find --quick). cm stats now breaks down scan usage by mode (--deep, --quick, and --diff).
• Pre-flight health checks (cm health): Added the cm health command (with --json, --strict, and --timeout flags) to verify credentials, configuration, endpoint reachability, and project authorization before running scans.
• Expanded file discovery: cm find now scans .hh, .hxx, .inl, and .ipp C++ headers, extensionless scripts with supported language shebangs (python, node, ruby, and php), nested build/, target/, obj/, and bin/ source directories, and production code under spec/, locale/, and i18n/.
• Bug fixes:
  • Restricted cm verify file writes to .exploit/ inside the project root and skipped missing or out-of-scope session artifacts during workspace restore.
  • Capped single-turn file reads at 256 KiB with pagination hints, bounded unified diff generation on large files, and raised run_command default and maximum timeouts to 120 seconds and 300 seconds.
  • Fixed an issue where sandboxed sessions could not write tool call entries to ~/.codemender/audit.log.
  • Updated multi-file --deep and --diff scans to report INCOMPLETE status when one or more file scans fail.

For more information, see the CodeMender documentation.

Feature

Supervised fine-tuning for Gemini 3 models is available in General Availability

Supervised fine-tuning for Gemini 3 models is available in General Availability (GA).

Key updates include the following:

• Gemini 3.5 Flash Lite support: Supervised fine-tuning is now supported for Gemini 3.5 Flash Lite (gemini-3.5-flash-lite), in addition to Gemini 3.5 Flash (gemini-3.5-flash) and Gemini 3.1 Flash Lite (gemini-3.1-flash-lite).
• Supported regions: Model tuning jobs and tuned model serving for Gemini 3 models are supported on the us and eu multi-region endpoints.
• Customer-managed encryption keys (CMEK): Customer-managed encryption keys (CMEK) are supported for protecting tuning job resources and tuned models.

For more information, see About supervised fine-tuning for Gemini models and Tune Gemini models by using supervised fine-tuning.

Feature

Priority PayGo support for Gemini Nano Banana 2.1

Priority Pay-as-you-go (PayGo) is available in General Availability (GA) for Gemini Nano Banana 2.1 (gemini-nano-banana-2.1) on the global endpoint.

Fixed

CodeMender updates (v0.15.0)

This release introduces updates to CodeMender:

• Enhanced SARIF 2.1.0 reporting: Updated cm report SARIF output with standardized rule descriptors, severity properties, and repository-relative artifact locations for improved GitHub Code Scanning compatibility.
• Model invocation metrics: Added model invocation tracking to cm stats --json, providing total, per-session, and per-turn invocation counts.
• Bug fixes:
  • Fixed an issue where token usage from completed turns was omitted from session logs during error or cancellation exit paths.

For more information, see the CodeMender documentation.

Feature

DeepSeek-V4.1-Flash is available in Public Preview

DeepSeek-V4.1-Flash from DeepSeek is available as a fully managed model (MaaS) in Model Garden. The model accepts text and image input, targets agentic and coding tasks, and supports a context window of up to 1M tokens.

For more information, see DeepSeek-V4.1-Flash.

Feature

Gemini 4 Argon in CodeMender (Preview)

CodeMender now supports Gemini 4 Argon. To use it, pass --model gemini-4-argon to cm find, cm verify, or cm fix. Gemini 3.8 Flash remains the default model.

For more information, see Specifying the model.

Feature

Gemini 4 Argon and Gemini 4 Argon Cyber Permissive

Gemini 4 Argon (gemini-4-argon) is available in General Availability (GA). Gemini 4 Argon is Google's frontier model that excels at deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science. It supports a unified 1,048,576-token input and output limit, configurable thinking levels (LOW, MEDIUM, HIGH), long decodes up to 1,048,576 tokens using continuation tokens on the GenerateContent API, and agentic video understanding in GA.

Gemini 4 Argon Cyber Permissive (gemini-4-argon-cyber-permissive), a version of Gemini 4 Argon for defensive cybersecurity use cases, is also available in GA behind the Fairwind Program allowlist.

• Known issue (long decodes): Long decoding (continuation_token) on the GenerateContent API (generateContent and streamGenerateContent) is supported only in the v1beta1 API (not v1). The Google Gen AI SDK uses v1beta1 by default unless api_version='v1' is specified in http_options; for direct REST calls, use the v1beta1 endpoint path.

For more information, see Get started with Gemini 4 and the developer's guide to Gemini 4 Argon.

Fixed

CodeMender updates (v0.14.0)

This release introduces updates to CodeMender:

• Better pull request scans: cm find --diff now catches removed security controls and gives large files more time (configurable with --diff-job-timeout).
• Unified sessions: --deep and --diff scans now appear as a single session in cm session list, cm report, and cm stats.
• Moved code in cm verify: cm verify now finds code that moved since cm find and updates the finding's location.
• Guided activation: cm now opens the activation page and waits for activation instead of exiting.
• New cm find flags: --deep-max-turns limits exploratory steps per file in --deep scans, and --enable-bash allows shell commands during scans.
• Improved detection: More distinct findings per file and more precise CWE classification for weak hashing and randomness.
• Bug fixes:
  • Fixed scans of repositories reached through nested symbolic links.
  • Fixed findings for nonexistent files and over-merged findings.
  • Shell scripts created by cm verify are now executable.

For more information, see the CodeMender documentation.

Feature

Anthropic's Claude Haiku 5.5

Claude Haiku 5.5 is available in Model Garden.

Fixed

CodeMender updates (v0.13.0)

This release introduces updates to CodeMender:

• Consistent finding deduplication: Improved finding deduplication consistency across repeated cm find, --deep, --diff, and --parallel scans.
• Bug fixes:
  • Fixed an issue where deduplicated findings could appear in cm report --status DISMISSED.
  • Improved local state reliability during concurrent cm find --diff and cm find --deep scans.
  • Updated streaming error handling so non-retryable errors fail immediately instead of retrying.
  • Ensured local session state and logs are saved cleanly when a session is interrupted (Ctrl+C).

For more information, see the CodeMender documentation.

Fixed

CodeMender updates (v0.12.0)

This release introduces updates to CodeMender:

• Hidden file scanning: Added the --include-hidden flag to cm find (or include_hidden under scan in config.yaml) to include hidden directories and files, such as .github/ and .config/, in vulnerability scans.
• Larger file support: Raised the default file discovery size limit (max_file_size_kb) from 500 KB to 2 MiB so cm find no longer skips large source files. Existing configurations that use the previous 500 KB default are upgraded automatically, and --deep and --diff scans can now read beyond the inline preview of files larger than 64 KB.
• Report output for every scan mode: cm find --output now writes the scan report for standard, --deep, --parallel, and --diff scans.
• Scoped fixes: cm fix now prevents the agent from modifying CI/CD pipelines, static analysis configurations, build definitions, and repository configuration files that are outside the scope of the vulnerability being fixed.
• Accurate session token usage: cm now reports the total token usage across all turns of a session, instead of only the final turn, so token counts in cm stats and session exports match billed usage.
• Bug fixes:
  • Fixed duplicate findings in cm report and cm stats, and stopped merging distinct vulnerability types reported on the same line.
  • Fixed database is locked errors during cm find --deep scans with multiple workers.
  • Reduced false positives by requiring a concrete, viable exploit before reporting a finding with high confidence.
  • Included sessions from cm find --deep and cm find --diff workers in cm stats.
  • Fixed cm verify resolving relative finding and exploit artifact paths against the current working directory instead of the project root.

For more information, see the CodeMender documentation.

Feature

Cloud Trace integration with Agent Gateway (Preview)

Agent Gateway integrates with Cloud Trace in

Preview

to provide end-to-end request observability for agent workloads. Enabling Cloud Trace gives you visibility into how requests travel from your agents through the gateway and across Google Cloud services, tools, agents, and MCP servers.

For more information, see

Use Cloud Trace

.

Feature

xAI's Grok 4.7

Grok 4.7

is available in

Preview

in Model Garden.

Fixed

CodeMender updates (v0.11.0)

This release introduces updates to CodeMender:

• Tiered location configuration: Added support for configuring the service location using the

CM_LOCATION

environment variable or

location

in

config.yaml

(defaulting to

global

), replacing the

--location

command-line flag.

• CI gate severity validation: Updated

cm find

to validate

--fail-on

severity values (

CRITICAL

,

HIGH

,

MEDIUM

,

LOW

, or

NONE

) up front instead of silently ignoring misspelled severities.

• Sandboxed VCS branching and hardening: Enabled

vcs_branch

creation and candidate branch cleanup inside the default sandboxed worker for

--architecture

sessions, and hardened VCS template expansion against shell metacharacters and command injection.

• Bug fixes:
  • Normalized vulnerability type casing and relative file paths during

cm find

and consensus runs to prevent duplicate findings.

• Extended HTTP 429 retry backoff (

Retry-After

support and up to 60-second quota refill windows) and automatic stream reconnection to improve scan resilience under rate limiting.

• Preserved standalone

cm find

,

cm verify

, and

cm fix

sessions on interrupt (

SIGINT

/

SIGTERM

) so interrupted runs can be resumed with

cm session resume

, while still canceling parallel consensus worker operations.

For more information, see the

CodeMender documentation

.

Feature

The App Topology API that provides agent topologies in Gemini Enterprise Agent Platform is now

generally available

This launch introduces the following changes:

• A new query builder for customizing suggested quick queries or creating your own query on the

Topologies

page. Custom queries let you explore more data related to your agents, such as identity, alerts, vulnerabilities, and underlying infrastructure. For details, see

View topologies for a project

.

• A streamlined experience for the

single-agent topology

in Agent Registry. A topology with single-hop traffic to and from the agent automatically loads on the page.

Feature

Custom denial messages for semantic governance policies

You can now configure a fixed Denial message on a semantic governance policy to show end users when the policy denies a request, in place of the policy denial rationale. The message can be up to 1,000 characters and is configurable in the Google Cloud console on the policy create or edit page, or by setting agentResponseCustomization.denialMessage on the policy resource using the REST API.

When a single request triggers denials from more than one policy, the policy engine combines the configured Denial messages, deduplicates them, and falls back to a generic message for firing policies that don't have one configured.

For more information, see Configure a semantic governance policy and Information exposure risk.

Feature

Anthropic's Claude Sonnet 5.5

Claude Sonnet 5.5 is available in Model Garden.

Feature

Provisioned Throughput: Support for changing order scope and increasing term

Provisioned Throughput now directly supports from the self service console the ability to change the scope or increase the term of an order through a supersede and replace operation.

For more information, see

Restructure an order
.

Feature

Gemini 3 models supported by the Interactions API in Preview

Gemini 3 models are supported by the Interactions API in Preview on Gemini Enterprise Agent Platform.

For more information, see the Interactions API overview
.

Feature

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS (Preview)

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are available in Preview on the global endpoint. The models support structured speaker and style metadata, inline vocal tags, multi-speaker dialogue, streaming, voice design, and voice replication.

For more information, see Generate speech with Gemini TTS
.