Google reports that Argon in its internal Antigravity environment verified 134 unresolved issues in under three hours: 97 stale or duplicate issues were closed with commit evidence, and the remaining 37 were routed by owner. The original post does not disclose its internal data sources, specific decision rules, or prompt. The workflow below is an editorial reproduction that requires human approval for any status change; it is not a step-by-step recreation of the internal process.
Good fit: Reviewing an issue backlog, with authorized access to code commit history, production release records, and relevant telemetry, to distinguish evidence-backed closure recommendations from issues that need an owner to handle them.
Poor fit: Automatically closing issues based only on merge records, a single query with no errors, or a model's judgment; claiming an issue is fixed without production or telemetry evidence.
Applicable model version: Gemini 4 Argon. The original post does not specify an API model ID, exact version, or reasoning level.
Applicable client, agent, or API: Google reports using Antigravity. The editorial reproduction requires an agent environment authorized to read the target repository, deployment records, and telemetry. The original post does not describe its internal connectors, tools, or access scope.
Recommended reasoning level and parameters: Not specified.
The original post does not disclose a reusable prompt. The task description below is an editorially prepared starting point for this workflow. Replace the fields and evidence sources to match the actual system; this does not represent Google's internal prompt.
Original Chinese copy:
逐条审阅我提供的未决缺陷。只使用当前获准访问的提交历史、生产部署记录和相关遥测。对每项分别列出缺陷 ID、关联提交、提交是否已部署到目标生产环境及其证据、部署后的遥测观察区间与结果、可能的重复项及其主记录。区分“有证据支持关闭建议”“建议交… This is a necessary excerpt; read the original source for full context.
English editorial adaptation:
Review each unresolved issue I provide. Use only the commit history, production deployment records, and relevant telemet… This is a necessary excerpt; read the original source for full context.
The steps below are an auditable reproduction method based on the original case, not an internal SOP disclosed by the source:
Fix the queue scope: Record the filters, query time, and total number of unresolved issues. Preserve each original issue ID; do not change the denominator during analysis.
Verify commit links: Find commits that may relate to each issue and record the evidence linking the commit ID to the issue. The existence of a commit or its merge shows only that a code change was made; it does not show that it was deployed or resolved the issue.
Verify production deployment: Check whether the relevant commit reached the target production environment and retain evidence of the deployment version, environment, and time. If you cannot verify the mapping between the commit and deployment version, mark the item “insufficient evidence.”
Check post-deployment telemetry: Look for the error signature described in the issue in the relevant period after deployment. Record the query scope, observation window, and results. If log coverage, sampling, or the time window is insufficient, do not treat “not observed” as “resolved.”
Make an item-by-item recommendation: Assign each item to exactly one review category: recommend closure with support from the commit, production deployment, and relevant telemetry; recommend marking it as a duplicate with a clear link to the primary issue; or request more evidence or refer it to the owner. Mark stale and duplicate items separately; do not infer the count in either category.
Update status after human review: The reviewer checks the evidence and recommendation for each item, then decides whether to close it, link it as a duplicate, or assign an owner. The agent does not change issue status directly. Record the reviewer, approval decision, execution time, and issue link; retain the original record and ownership.
Reconcile the counts: If the queue is divided into closure recommendations and owner referrals, check initial unresolved count = closure count + referrals + items still awaiting review. The pending-review count is zero only when every item has been classified. If reporting subtotals such as duplicate or insufficient-evidence items, state whether they are subdivisions of closure recommendations or referrals to avoid double-counting.
Google's Gemini 4 Argon in Google Antigravity was published on 2026-10-13. The page describes the case as an internal Alphabet engineering practice: Argon orchestrated agents to verify 134 long-standing issues across multiple services, cross-checking commit history, whether candidate fixes had reached production, and whether errors were still triggering in live telemetry.
The same page reports that, in under three hours, 97 stale or duplicate issues were closed with commit evidence, while another 37 issues were categorized by owner. 97 + 37 = 134; 97 / 134 ≈ 72.4%, the queue-reduction rate reported in the original post. The post does not break down how many of the 97 were stale versus duplicates, and it does not publish an itemized list.
The counts, elapsed time, and outcome above are from Google's internal case report, not an independent test. This workflow does not claim to have rerun the case or treat those figures as expected performance for other teams.
The original post does not disclose the repository, issue tracker, data-source connectors, commit-matching method, production deployment criteria, telemetry query, or observation window. It also does not explain how human review was handled during the run. The fields, classifications, and human-approval threshold in this article are therefore editorial reproduction recommendations.
Anyone reproducing the workflow must already be authorized to access the relevant repository, production deployment records, and telemetry. This process does not provide a way to bypass login, access controls, or permission approvals. If evidence is unavailable, leave the item pending or refer it to the owner.
Issue status changes affect team records. Model output can only serve as a sourced recommendation; authorized people must review and carry out final closures, duplicate links, and assignments.
97/134 is the backlog ratio in this case, not model accuracy. The under-three-hour duration is the reported time for this internal run and is not guaranteed to recur with other backlogs, systems, or permission setups.
Use an organization-approved Antigravity or other authorized agent environment. Prepare unresolved issues with a clearly defined query time, along with code commits, target production deployment records, and relevant telemetry that can be cross-checked. First generate an itemized evidence table and recommendations in read-only mode; have an owner review them before manually updating issues. Record the queue scope, counts by category, unverifiable items, and review change log. Do not assume the result will be 97/134 or that the work will finish within three hours. The original post does not provide enough input data or runtime configuration to independently reproduce its internal result.
Gemini 4 Argon