{"slug": "vs-code-1-137-puts-ai-agents-in-dev-containers-2026", "title": "VS Code 1.137 Puts AI Agents in Dev Containers [2026]", "summary": "Microsoft released Visual Studio Code 1.137 on September 9, 2026, introducing an Agent Host system that runs multiple AI coding agents as managed processes inside the editor, each in its own isolated Dev Container with its own tools and dependencies. The release, followed by an update roughly a week later, also adds shared session access so a developer or a second agent can inspect and interact with a running agent session, plus new automations and a voice mode. The container boundary contains damage from destructive commands or broken installs to the Dev Container rather than the developer's filesystem, as the update landed alongside Google's public developer access to Gemini 4 Argon, OpenAI's GPT-6 Astra passing 87% programming-language coverage, and a DeepSeek coding-focused model.", "body_md": "Microsoft has quietly turned its most popular code editor into something closer to an operating system for AI agents. Visual Studio Code 1.137, released September 9, 2026, and the follow-up update that landed roughly a week later, add a feature Microsoft calls the Agent Host system: an architecture that lets multiple AI coding agents run as managed processes inside the editor, each with its own isolated Dev Container, its own tools, and its own dependencies. Paired with new automations and a voice mode, the release marks the clearest sign yet that Microsoft sees the editor’s future as a control center for autonomous software work, not just a faster autocomplete box.\n\nThe timing matters. The update landed in the same month Google opened Gemini 4 Argon to public developer access, OpenAI pushed GPT-6 Astra’s programming-language coverage past 87%, and DeepSeek shipped a coding-focused model built for end-to-end project work. VS Code’s agent infrastructure isn’t arriving in a vacuum. It’s arriving right as the underlying models get cheap and capable enough to actually run unattended for long stretches, which is the whole point of giving them a container to live in.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## What the Agent Host System Actually Does\n\nAccording to the [official VS Code 1.137 release notes](https://code.visualstudio.com/updates/v1_137), the update reworks how the editor manages AI-driven processes. Instead of treating an AI coding assistant as a single chat panel bolted onto the sidebar, the Agent Host system gives each agent session its own managed runtime. That runtime can be spun up, paused, monitored, and torn down independently of the editor’s main process, which is the kind of plumbing you’d expect from a container orchestrator, not a text editor.\n\nDeveloper coverage of the release, including a [write-up on DEV Community](https://dev.to/hamidrazadev/vs-code-1137-is-here-automations-voice-mode-and-the-features-thatll-change-how-you-code-lgg), frames 1.137 as pushing the editor “closer to a real AI-powered coding partner” rather than a passive tool. The practical upshot: a developer can now kick off a background agent to refactor a module, keep working in the foreground on something else, and check back on the agent’s progress without losing their own editor state.\n\n## Agents Now Run Inside Dev Containers\n\nThe headline technical change is that agents can run inside local Dev Containers, each preloaded with the project’s own tools and dependencies. That’s a meaningful shift from running an agent directly against a developer’s host machine. A Dev Container gives the agent a sandboxed, reproducible environment that matches the project’s actual build and test setup, rather than whatever happens to be installed globally on the developer’s laptop.\n\nIndustry analysis published around the update, including a piece from [Analytics Insight](https://www.analyticsinsight.net/artificial-intelligence/how-ai-is-changing-code-editors-in-2026), describes this as part of a broader pattern: “code editors become agent workspaces” where “one editor can handle many agents” at once. That framing lines up with what Microsoft shipped. The container boundary isn’t just a convenience, it’s a safety mechanism. If an agent runs a destructive command or installs something broken, the damage is contained to that Dev Container instead of spreading to the developer’s actual filesystem.\n\n### Shared Session Access\n\nThe update also adds shared session access, letting a developer or a second agent inspect and interact with a running agent session instead of starting fresh each time. That matters for teams: a senior engineer can look at what a background agent is doing mid-task, nudge it in a different direction, or hand off a half-finished task to a colleague without re-explaining the entire context from scratch.\n\n### Codex Support Baked In\n\nReports on the release also note added Codex support, meaning OpenAI’s coding tooling gets a more native hook into the Agent Host system alongside Microsoft’s own Copilot stack. That’s consistent with how Microsoft has approached AI tooling in VS Code generally: the editor stays model-agnostic at the infrastructure layer, even while Copilot remains the default experience most developers see first.\n\n## Automations and Voice Mode Round Out the Release\n\nAlongside the Agent Host work, VS Code 1.137 ships automations, which let developers define trigger-based actions tied to editor or repository events, and a voice mode for hands-free interaction with Copilot and agent sessions. Neither feature is as structurally significant as the Dev Container work, but together they point in the same direction: VS Code is adding the scaffolding of a workflow-automation platform on top of what used to be a text editor with extensions.\n\nVoice mode in particular signals Microsoft’s bet that typing won’t remain the only interface for directing an AI coding agent. If an agent is already running semi-autonomously inside a container, dictating a correction or a new instruction out loud is arguably a more natural interaction than switching back to a keyboard-driven chat box.\n\n## What Changed in VS Code 1.137, at a Glance\n\n| Feature | What It Does | Shipped | Why It Matters | \n|---|---|---|---|\n| Agent Host system | Runs AI agents as managed, isolated processes inside the editor | September 2026 release cycle | Core infrastructure for running multiple agents at once | \n| Dev Container agent execution | Runs agents inside local Dev Containers with the project’s own tools and dependencies | September 2026 follow-up update | Sandboxes agent actions away from the host machine | \n| Shared session access | Lets a developer or second agent inspect and join a running agent session | September 2026 release cycle | Supports handoffs and mid-task collaboration | \n| Automations | Trigger-based actions tied to editor and repository events | VS Code 1.137, September 9, 2026 | Moves the editor toward workflow automation, not just editing | \n| Voice mode | Voice-driven interaction with Copilot and agent sessions | VS Code 1.137, September 9, 2026 | Adds a hands-free interface for directing agents | \n| Codex support | Native hooks for OpenAI’s Codex tooling inside the Agent Host | September 2026 release cycle | Keeps the editor’s agent layer model-agnostic | \n\n## Why Microsoft Is Betting on Agents Over Autocomplete\n\nFor most of the last three years, AI in code editors meant one thing: inline autocomplete that finishes a line or a function while you type. That model is cheap to run and easy to trust, because a developer reviews every suggestion before accepting it. Agent-based coding is a different bet entirely. An agent is handed a task description, goes off and works on it across multiple files for minutes or hours, and comes back with a diff. The developer’s role shifts from “accept or reject this line” to “review this entire change before it merges.”\n\nReports on Gemini 4 Argon’s launch describe it as part of a broader shift toward coding agents that manage multi-step software tasks rather than only autocomplete. VS Code’s Agent Host system is the editor-side infrastructure that shift needs. Without a safe place to let an agent run unattended, autocomplete remains the ceiling. With Dev Container isolation, Microsoft is betting developers will trust agents with bigger, longer-running jobs.\n\n## The Model Wave Feeding These Agents\n\nNone of this infrastructure matters without models capable of filling it. October 2026 has brought an unusually dense run of releases aimed specifically at agentic coding. Google opened Gemini 4 Argon to the Gemini API and Google AI Studio on October 13, under the stable API identifier gemini-4-argon, positioning it for complex, long-running software-development workflows. Google AI Ultra subscribers can also reach Argon through Google Antigravity, an agent-oriented development environment, and reports say the model is already being integrated into third-party coding tools including Cursor.\n\nPricing for Argon’s API access is set at $2 per million input tokens and $10 per million output tokens through November 30, 2026, after which it rises to $4 and $20 respectively, according to the same reporting. That kind of promotional pricing window is a familiar pattern for frontier labs trying to seed adoption before a model’s real cost structure kicks in.\n\nOpenAI, meanwhile, has pushed GPT-6 Astra’s language coverage well past its predecessor: Astra reportedly reaches an 87.4% average pass rate on MultiLangBench across 20 programming languages, compared with 26.8% for GPT-5.6 Sol, with particular gains in lower-resource languages like COBOL, Fortran, and Visual Basic. DeepSeek released V4.1-Flash-1017 on October 17, describing it as a reinforcement-learning-powered coding model built for end-to-end project work, including game development and 3D modeling tasks. And MiniMax’s general release of M3.1 Flash on October 19 targets the cheap-and-fast end of the market, priced at $0.60 per million output tokens for coding, computer use, and long agent tasks.\n\nThat price compression is arguably as important as any single model’s benchmark score. An agent that stays resident in a Dev Container for an hour-long task burns far more tokens than a single autocomplete suggestion ever would. Editors like VS Code can only make agentic workflows the default once running an agent for that long stops being prohibitively expensive.\n\n## Competitive Landscape: VS Code vs. Cursor vs. GitHub Copilot\n\nVS Code’s Agent Host push doesn’t happen in isolation from the rest of the editor market. [GitHub Copilot recently retired six legacy models and shifted its default to GPT-5.6](https://shattered.io/github-copilot-retires-6-models-gpt-5-6-2026/), a reminder that Microsoft is managing two parallel AI surfaces inside the same ecosystem: Copilot as the integrated assistant, and the Agent Host as the underlying infrastructure that any model or tool can plug into.\n\nCursor, built by Anysphere, has taken a different approach by designing its entire editor around agent-first workflows from the start, and has already added Gemini 4 Argon support following the model’s October 13 launch. Where VS Code is retrofitting agent infrastructure onto an editor with a 14-year install base and a massive extension ecosystem, Cursor built agent-centric design in from day one. The tradeoff cuts both ways: VS Code brings reach and compatibility with Microsoft’s broader dev tooling, including [AI agent controls Google has been adding to Chrome DevTools](https://shattered.io/chrome-devtools-154-ai-agent-controls-2026/), while Cursor brings a more opinionated, purpose-built agent experience.\n\nAnthropic’s Claude Code line has moved in a similar direction, with recent updates turning on fork subagents by default across its command-line tool, Agent SDK, desktop app, and VS Code extension. That convergence, four separate players all building toward the same multi-agent, containerized model, suggests the editor market has settled on roughly the same architecture even while competing on model choice and price.\n\n## Agentic Coding Tools at a Glance, October 2026\n\n| Tool | Maker | What Changed in October 2026 | Pricing Signal | \n|---|---|---|---|\n| VS Code (Agent Host) | Microsoft | Agents now run inside isolated Dev Containers; automations and voice mode added | Editor is free; backing model costs billed separately | \n| GitHub Copilot | Microsoft / GitHub | Retired six legacy models, shifted default to GPT-5.6 | Subscription tiers | \n| Cursor | Anysphere | Added Gemini 4 Argon integration | Subscription plus usage-based billing | \n| Gemini 4 Argon (API) |  | Opened to public developer access on October 13 | $2/M input, $10/M output through Nov 30; $4/$20 after | \n| DeepSeek V4.1-Flash-1017 | DeepSeek | Released October 17 for end-to-end coding and project work | Not disclosed in available reporting | \n| MiniMax M3.1 Flash | MiniMax | General release on October 19 for agentic coding tasks | $0.60/M output tokens | \n\n## The Developer Adoption Numbers Behind the Shift\n\nThe push toward agentic editors tracks with how developers say they’re already working. Stack Overflow’s 2025 Developer Survey found that “84% of developers reported that they use or plan to use AI tools in their development process,” according to the [2025 Stack Overflow Developer Survey](https://survey.stackoverflow.co/2025/technology). More specifically, Stack Overflow reported that “31% of developers reported that they were currently using AI agents in 2025,” per its own [breakdown of the survey results](https://stackoverflow.blog/2025/08/01/diving-into-the-results-of-the-2025-developer-survey/), a figure that’s almost certainly higher a year later given how many agent-capable tools have shipped since.\n\nThe same survey data shows 51% of professional developers use AI tools daily, and 81.4% reported using OpenAI’s GPT models for development work over the prior year. On tool-specific usage, Stack Overflow’s survey found 18% of developers regularly use Cursor, 10% use Claude Code, and 5% use Windsurf among AI-enabled development environments, numbers that give some sense of how fragmented the agentic-editor market remains even as VS Code pushes new infrastructure into the mainstream Copilot user base.\n\n## Historical Context: From IntelliSense to Autonomous Agents\n\nVS Code’s own history makes the scale of this shift easier to see. The editor launched in 2015 built around IntelliSense, a static, rules-based code-completion engine with no AI involved at all. [GitHub Copilot](https://shattered.io/benchmark-cursor-vs-github-copilot-2026/)‘s 2021 debut added the first mainstream AI-powered suggestions, but those suggestions were still line-by-line and required constant human review. Chat-based assistants followed, letting developers ask questions about their codebase without leaving the editor, but the developer still drove every action.\n\nThe Agent Host system represents a third phase: the AI no longer waits for a prompt to generate one suggestion, it runs semi-independently across a task that can span many files and many minutes. [Command-line agent tools like Claude Code, Codex CLI, and OpenCode have been iterating on this model for months](https://shattered.io/claude-code-codex-cli-opencode-update-4-days-2026/), but VS Code bringing the same architecture into a GUI editor used by millions of developers is a different kind of milestone. It’s the difference between a niche workflow and a default one.\n\n## Market Impact for Microsoft, GitHub and Rivals\n\nFor Microsoft, embedding agent infrastructure directly into VS Code rather than keeping it exclusive to Copilot is a notable strategic choice. It means rival models, including Gemini 4 Argon and OpenAI’s Codex tooling, can run inside the same Agent Host plumbing that powers Copilot. That’s either a confident bet that Copilot wins on integration and distribution even in an open playing field, or a recognition that trying to lock developers into one model would backfire given how fast the underlying model landscape is moving.\n\nFor [GitHub Copilot](https://shattered.io/cursor-rules-migrate-github-copilot-2026/) specifically, the Agent Host update arrives just after Copilot’s own model lineup churned, with six legacy models retired in favor of GPT-5.6 as the default. Pairing a more current model lineup with deeper agent infrastructure is a defensive move against Cursor, which has built its entire pitch around being the fastest adopter of whichever frontier model performs best on coding benchmarks that week.\n\nThe broader AI tooling market is also reacting to the same pressure. [Gemini 4 Argon’s rollout](https://shattered.io/gemini-4-argon-global-launch-16x-tokens-2026/) and [DeepSeek’s V4.1-Flash-1017 release](https://shattered.io/deepseek-v4-1-flash-1017-hle-score-42-6-2026/) both landed within the same week as VS Code’s agent infrastructure matured, which suggests editor vendors and model labs are now moving on roughly synchronized release cycles, each racing to make sure their product is ready the moment the other side ships something better.\n\n## Friction Points: Security, Container Sprawl and Review Fatigue\n\nNone of this is friction-free. Running agents inside Dev Containers solves one security problem, host-machine isolation, while potentially creating others. Every active agent session now needs its own container, and teams running several agents across several projects will need to manage container sprawl: images to maintain, resource limits to tune, and credentials to scope correctly so an agent in one container can’t reach secrets meant for another.\n\nThere’s also a human bottleneck that container isolation doesn’t fix: review fatigue. If an agent can now work unattended for an hour instead of ten seconds, it also produces a correspondingly larger diff for a human to review before merging. [AI-assisted code review tools have been racing to keep pace with this exact problem](https://shattered.io/ai-code-review-tools-bugbot-vs-codex-2026/), but the fundamental tension remains: letting agents work longer without supervision is only safe if someone, human or automated, is checking the output just as carefully as before.\n\n## Five Predictions for Agentic Coding Through 2027\n\n- Expect other major editors, including JetBrains’ IDE line and Zed, to announce their own container-isolated agent runtimes within the next two to three release cycles, following the pattern VS Code just set.\n- Model pricing for agentic coding will likely keep compressing, following the trajectory set by Gemini 4 Argon’s promotional window and MiniMax M3.1 Flash’s $0.60-per-million-token rate, as labs compete for the high-token-volume agent workload rather than one-off chat queries.\n- Code review tooling will become the next major battleground, since longer-running agents produce larger diffs; expect more vendors to pitch AI-assisted review as a required companion product to AI-assisted coding, not an optional add-on.\n- Container and sandbox management for AI agents will likely spin off into its own tooling category, separate from general-purpose Dev Containers, as teams need finer-grained controls over what an autonomous agent can access.\n- Expect [GitHub Copilot](https://shattered.io/github-copilot-vs-cursor-4-models-8-days-2026/) and Cursor to keep leapfrogging each other on model support in the near term, with each new frontier release (Gemini 4 Argon, GPT-6 Astra, and whatever follows) triggering a fast integration race rather than exclusive partnerships.\n\n## What Developers Should Do Right Now\n\nDevelopers curious about the Agent Host system don’t need to wait for a broad rollout; the features shipped in the September 2026 release cycle and are available in current VS Code builds. The more practical question is whether a given team’s project is already set up with a Dev Container configuration. Agents running inside containers only get the benefit of project-specific tools and dependencies if that container definition already exists and is kept current, which means teams without a working devcontainer.json are better served by setting that up first before leaning on agent automation inside it.\n\nTeams evaluating whether to lean on VS Code’s native agent tooling versus a dedicated agent-first editor like Cursor, or a command-line tool, should weigh existing workflow investment heavily. An organization with deep VS Code extension dependencies and GitHub-centric CI/CD has less to gain from switching editors than from adopting the Agent Host features already built into the tool they use today.\n\n## Frequently Asked Questions\n\n**What is the VS Code Agent Host system?**\n\nIt’s infrastructure added in the September 2026 VS Code release cycle that runs AI coding agents as managed, isolated processes inside the editor, with support for running those agents inside local Dev Containers.\n\n**When was VS Code 1.137 released?**\n\nVS Code 1.137 was released on September 9, 2026, adding automations and voice mode. The Dev Container agent execution and Agent Host capabilities followed in the same release cycle.\n\n**Do I need Dev Containers set up to use VS Code’s AI agents?**\n\nTo get the isolation and project-specific tooling benefits, yes. Agents can still run without a Dev Container, but running them inside one keeps their actions sandboxed away from your host machine and matched to your project’s actual build environment.\n\n**Does VS Code’s Agent Host work with models other than GitHub Copilot?**\n\nReports indicate the Agent Host system includes support for OpenAI’s Codex tooling, and the broader AI coding model landscape, including Gemini 4 Argon, is expanding quickly enough that third-party integrations are likely to keep growing.\n\n**How does VS Code’s agent system compare to Cursor?**\n\nCursor was designed around agent-first workflows from the start, while VS Code is adding agent infrastructure, including Dev Container isolation, on top of its existing editor and extension ecosystem. Both now support multiple frontier models, including Gemini 4 Argon.\n\n**Is voice mode available for all VS Code AI features?**\n\nVoice mode shipped in VS Code 1.137 for interacting with Copilot and agent sessions. Available reporting doesn’t specify every feature it currently covers.\n\n**What percentage of developers already use AI agents?**\n\nStack Overflow’s 2025 Developer Survey found 31% of developers reported using AI agents, while 84% said they use or plan to use AI tools in their development process generally.\n\n**Will running agents in Dev Containers slow down my machine?**\n\nRunning multiple containerized agents does add resource overhead compared to a single lightweight autocomplete suggestion, since each agent session needs its own container. Available reporting doesn’t include specific performance benchmarks for the Agent Host system’s resource usage.", "url": "https://wpnews.pro/news/vs-code-1-137-puts-ai-agents-in-dev-containers-2026", "canonical_source": "https://shattered.io/vs-code-1-137-agent-host-dev-containers-2026/", "published_at": "2026-10-20 12:29:33+00:00", "updated_at": "2026-10-20 12:49:15.354354+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "ai-tools", "ai-products"], "entities": ["Microsoft", "Visual Studio Code 1.137", "Agent Host", "Dev Containers", "Google", "Gemini 4 Argon", "OpenAI", "GPT-6 Astra"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/vs-code-1-137-puts-ai-agents-in-dev-containers-2026", "markdown": "https://wpnews.pro/news/vs-code-1-137-puts-ai-agents-in-dev-containers-2026.md", "text": "https://wpnews.pro/news/vs-code-1-137-puts-ai-agents-in-dev-containers-2026.txt", "jsonld": "https://wpnews.pro/news/vs-code-1-137-puts-ai-agents-in-dev-containers-2026.jsonld"}}