---
type: "post"
id: "codemender-overview-gemini-enterprise-agent-platform-b873d6"
slug: "codemender-overview-gemini-enterprise-agent-platform-b873d6"
title: "CodeMender overview: Gemini Enterprise Agent Platform"
description: "Google Cloud’s CodeMender documentation lists Gemini 4 Carbon as a supported model for its AI code-security agent, alongside Gemini 4 Argon and earlier Gemini models."
retrieval_nugget: "Google Cloud’s CodeMender documentation lists Gemini 4 Carbon as a supported model for its AI code-security agent, alongside Gemini 4 Argon and earlier Gemini models. Teams evaluating automated vulnerability work can now select Carbon for find, verify, and fix commands, but CodeMender remains a limited Public Preview for testing and evaluation—not production use."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["gemini","google"]
entities: ["Google Cloud","CodeMender","Gemini 4 Carbon","Gemini 4 Argon","Gemini 3.8 Flash","Gemini Enterprise Agent Platform","Google DeepMind"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
source_url: "https://docs.cloud.google.com/gemini-enterprise-agent-platform/agents/codemender"
source_urls: ["https://docs.cloud.google.com/gemini-enterprise-agent-platform/agents/codemender"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:codemender-overview-gemini-enterprise-agent-platform-b873d6"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/codemender-overview-gemini-enterprise-agent-platform-b873d6/","markdown":"https://newruntime.com/posts/codemender-overview-gemini-enterprise-agent-platform-b873d6.md","json":"https://newruntime.com/posts/codemender-overview-gemini-enterprise-agent-platform-b873d6.json"}
markdown_url: "https://newruntime.com/posts/codemender-overview-gemini-enterprise-agent-platform-b873d6.md"
---

# CodeMender overview: Gemini Enterprise Agent Platform

## Retrieval answer

Google Cloud’s CodeMender documentation lists Gemini 4 Carbon as a supported model for its AI code-security agent, alongside Gemini 4 Argon and earlier Gemini models. Teams evaluating automated vulnerability work can now select Carbon for find, verify, and fix commands, but CodeMender remains a limited Public Preview for testing and evaluation—not production use.

## Summary
Google Cloud’s CodeMender documentation lists Gemini 4 Carbon as a supported model for its AI code-security agent, alongside Gemini 4 Argon and earlier Gemini models. Teams evaluating automated vulnerability work can now select Carbon for find, verify, and fix commands, but CodeMender remains a limited Public Preview for testing and evaluation—not production use.

## What happened
Google Cloud describes CodeMender as an AI code-security agent that finds vulnerabilities, verifies them, and generates patches. It wraps a fine-tuned harness around a large language model, using Google DeepMind-engineered prompts, skills, and orchestration. The documented workflow can scan a codebase for vulnerabilities or import findings from external security scanners; build code and attempt exploits to check whether findings are exploitable; and generate and test a language-compatible patch. Users can provide application and threat-model context throughout those stages.

The overview lists Gemini 4 Carbon as a supported model, along with Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash, and Gemini 3.1 Pro Preview. Gemini 3.8 Flash is the default. To choose Carbon, users pass `--model gemini-4-carbon` to `cm find`, `cm verify`, or `cm fix`. The page also lists a Carbon cyber variant, `gemini-4-carbon-cyber-permissive`, for customers with Fairwind Program access. Google says safety blocks on Gemini 4 Argon or Carbon automatically trigger a retry with Gemini 3.8 Flash Cyber, then Gemini 3.8 Flash if the project lacks access to the cyber model.

CodeMender has two components: a hosted multi-agent system for its core reasoning and business logic, and a local client that acts as a CLI and daemon. The client executes commands on the user’s machine and can use a local process-level sandbox for compiling code, running tests, and verifying vulnerabilities. By default, commands are restricted to the workspace and command policy. Google warns that disabling the sandbox or bypassing path or command boundaries calls for running the CLI in an isolated VM or container.

The default language set covers C/C++, C#/.NET, Go, Java, JavaScript and TypeScript, Kotlin, Python, Ruby, Rust, and PHP, with common enterprise frameworks including Django, Flask, React, Spring Boot, ASP.NET, and Express. Additional file extensions can be included globally or per repository in `config.yaml`. Google says the listed languages are not a hard limit, but reports no formal per-language evaluations and notes results outside the benchmarked default languages will vary. CodeMender is globally available only to a limited set of customers in Public Preview; Google directs interested customers to contact their sales team. The terms limit use to testing and evaluation, prohibit commercial or production use, and restrict analysis to owned or authorized code or OSI-licensed open-source code for legitimate defensive purposes.

## What the archive adds
New Runtime covered the Gemini Enterprise Agent Platform’s model lifecycle on [November 13, 2026](https://newruntime.com/posts/model-versions-and-lifecycle-gemini-enterprise-agent-platform-6962a5/), noting Gemini 4 Hydrogen alongside Carbon. This is a new fact on that platform story: the CodeMender overview now names Carbon as a selectable model for security-agent tasks, rather than merely documenting the platform’s model roster.

## Trend signal
This supports the view that production agent systems need a durable harness around interchangeable models, rather than depending on one model identity. CodeMender combines a specialized harness, tools, and orchestration with a model choice exposed through CLI commands. See [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). This is an implementation signal, not evidence that teams have adopted the approach at scale.

## Why now
Google’s documentation was last updated on November 13, 2026, and includes Carbon among supported models. The page frames the product against rising concern about code vulnerabilities and attackers’ ability to generate exploits, but does not tie the model listing to a separate launch announcement or explain why this change happened at this time.

## Why it matters
Model selection is only one part of an automated security workflow. CodeMender’s documented process connects finding to verification and patch testing, while its local client performs builds, tests, and command execution in the user’s environment. That separation gives security teams a concrete evaluation target: not simply whether a model can suggest a fix, but whether the system can produce a finding, test exploitability, and validate a proposed patch under controlled conditions.

The constraints are equally consequential. CodeMender is still preview-only and explicitly barred from production use. Its execution can modify files, so sandbox boundaries and isolation are operational requirements, not optional polish. Teams considering an evaluation should treat Carbon as one model option inside a supervised workflow, define what code may be analyzed, and test results against their own threat models before making any deployment decision.

## Limits and uncertainty
Google says CodeMender is available to a limited set of customers in Public Preview and offers no formal per-language evaluations. The documentation does not establish how Carbon performs relative to the default Gemini 3.8 Flash on vulnerability discovery, exploit verification, patch correctness, or false-positive reduction. It also does not provide an independent evaluation of those outcomes. Preview terms prohibit production and commercial use, so the documented capabilities should not be read as evidence of production readiness.

## Reader takeaway
Gemini 4 Carbon is now a selectable CodeMender model, but the meaningful unit to evaluate is the whole supervised workflow: finding, verification, and tested patching. For now, that evaluation belongs in a properly isolated, non-production preview environment—not in a live security pipeline.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
