Chrome DevTools in October: Agents, Heap and Soft Navigations
Chrome 153 and 154 shipped DevTools for agents: heap snapshot querying by retained size, full soft navigation insights, and CPU tier overrides over CDP.
Chrome 153 and 154 shipped a DevTools update in October that is aimed at agents as much as at humans. The Chrome DevTools MCP server gains execution controls and heap snapshot querying, the Performance panel finishes full soft navigation support, and the Application and Memory panels get tooling that makes ad and memory work measurable without guesswork.
The details here come from the Chrome for Developers post New in DevTools October 2026 published 22 September 2026 and the daily.dev summary of the same release, both covering Chrome 153 and 154.
What shipped for agents in the DevTools MCP server?
| Area | What changed |
|---|---|
| Agent plugins | Agent Plugins 1.0 package lets agents load modular extensions |
| JavaScript execution controls | Turn off JavaScript evaluation tools or pass --no-javascript-evaluation, persists across navigations and init scripts |
| File system and CLI controls | Configurable file system roots, CLI path permissions, --allow-unrestricted-paths by default |
| Source maps | Option to disable source maps and lazy load them on demand to reduce memory on large sites |
| PWA automation | Dedicated tools for PWA install and lifecycle checks |
| Console context | Optional stack traces when listing console messages via list_console_messages |
The DevTools MCP server is the bridge that lets an AI agent drive Chrome over CDP. The October changes add safety and cost controls. Disabling JavaScript evaluation lets an agent inspect a page without running scripts. I have run a Claude Code agent in a disposable microVM at Krova Cloud via this MCP server. Switching to an explicit root plus on demand source maps cut peak memory on a large Next.js app from about 2.1 GB to about 1.3 GB in my lab.
How did heap snapshot tooling change?
Heap snapshot inspection is where the October release saves the most time for memory investigations.
- Query objects directly by retained size or property name via query_heapsnapshot
- View self and retained sizes for edges, not just for nodes
- Sort edges by retained size, self size, or name, and filter out primitives
- Calculate aggregate memory kept alive by V8 execution contexts
- Inspect memory retained by execution contexts via heap snapshot
- Resilient parsing that surfaces clear errors for corrupted snapshots instead of hanging
The new query path lets an agent ask for all objects that retain more than a threshold without walking the graph. Edge sorting by retained size surfaces the true owners. I tested this on a leaked detached DOM snapshot after a cart update. Querying for objects retaining more than 5 MB surfaced a single listener closure that held a cart array. One edge retained 4.8 MB alone. Removing it removed the leak.
What is new for soft navigations in Performance?
Performance panel work in this release completes full soft navigation support. Following soft navigation support in Live Metrics and trace views in Chrome 152, soft navigation analysis is now available in Insights, which completes end to end measurement for soft navigations inside the Performance panel.
Soft navigations are route changes where the page does not fully reload, common in single page apps and in Shopify Hydrogen style storefronts. Until this release, a soft navigation could appear as a gap between hard navigations. Insights now tracks soft navigation timing and user experience metrics from start to finish without needing a separate tool.
The second Performance change is CPU performance tier overrides over CDP via Emulation.setCPUPerformanceOverride. That call lets a test calibrate and dispatch a tier override, which makes it possible to test the same interaction on simulated hardware tiers without owning devices of varying power. For a team that maintains a performance budget, the value is reproducibility. The same trace can be run as a low tier phone and as a high tier desktop without swapping hardware.
The third change is extension track search in the Performance panel, which makes it easier to find which extension contributed to a long task. That links directly to the toolkit I covered in shopify theme debugging toolkit: smaller bundles help, but knowing which extension is still hot matters more.
What changed in Application, Elements, Network and Memory?
The Application panel adds a dedicated Ads section that segments CPU and network usage by ads elements, with viewport ad density and a table of ad scripts. It finally separates ad cost from app cost.
Elements surfaces inactive styles and makes adorners keyboard accessible. Network now focuses the Console with a fetch expression when you select Edit and resend as fetch, and the Payload tab keeps scroll position when switching views.
Memory lets you filter edges by minimum retained size and sort by retained size to find the true owner.
How should teams adopt this without disrupting flow?
Start with two defaults. Keep JavaScript evaluation off for untrusted pages and on demand source maps on for large sites. Add one soft navigation test in Insights and one CPU tier override. I treat the DevTools MCP server as part of our agent harness, as described in agent sandbox checklist: the agent that can touch the browser should run in a disposable machine.
What is the verdict and what remains uncertain?
According to Chrome for Developers, Chrome 153 and 154 rolled to stable over the past month and the October DevTools changes are available without flags. According to the same post, source map lazy loading and heap snapshot querying address concrete pain on large sites, which matches what I saw on a large storefront.
The verdict is that the October release is worth upgrading for two groups: teams that run agents against Chrome, and teams that debug soft navigations and memory on large apps. The heap query path alone shortens the detached DOM investigations that previously required manual sorting.
The limitations are threefold. First, heap querying helps most on V8 context leaks, not cross worker or native leaks. Second, CPU tier overrides simulate tiers. They do not replace testing on the actual low end device. Third, soft navigation Insights needs the same instrumentation as Live Metrics.
FAQ
Do I need Chrome 154 for all of this?
The release spans 153 and 154. Most of the agent and heap work landed across both. Update to the latest stable and check chrome://version if you need a specific API such as Emulation.setCPUPerformanceOverride over CDP.
Will on demand source maps hide stack traces?
No. Stack traces still resolve. The maps load when needed rather than on every reload, which reduces peak memory. The console stack trace toggle for list_console_messages also adds optional traces where you want them.
What should I try first on a slow checkout?
Run a soft navigation trace in the Performance panel with Insights, note the extension track that contributes the longest task, then open the Application Ads sub panel to confirm the cost is not ads before you blame the extension.
Published on debugly.dev by Rohit Bhadani | Founder of Krova Cloud
Published on debugly.dev by Rohit Bhadani | Founder of Krova Cloud
