VS Code Large Agent Session Lists: PR Extension Analysis

Your editor used to worry about tabs. Now it worries about crowds of semi-autonomous interns with commit access. The useful signal in Tech Insider’s new walkthrough is not simply that someone can build a VS Code extension for agent-generated pull requests. It is that the editor is starting to look like an air traffic control tower for bots that open PRs, leave artifacts, and then vanish like they have a lasagna in the oven.

The release note became the roadmap

According to Tech Insider’s Nadia Dubois, VS Code shipped version 1.139.1 on September 25, 2026, and its release note says the editor needs to handle "large agent session lists" faster. Tech Insider also reports that three days earlier, Cursor launched two bots, Rollouts and Security Review, which watch pull requests through deployment and flag exploitable bugs automatically. Put those together and the product signal gets loud: agent coding is no longer just autocomplete wearing a trench coat. It creates sessions, PR state, review artifacts, and security findings that developers need to see without spelunking through six tabs like a raccoon in a data center.

That is why Tech Insider’s 14-step tutorial matters beyond the mechanics of building a VS Code extension. The gap it describes is a missing unified place to see agent sessions, PR automation, and security scanning, especially when a team is running multiple autonomous coding agents at once. The editor becomes the coordination layer not because IDEs are sacred temples, but because developers already live there. If the robots are filing paperwork, somebody has to build the desk.

Agent PRs stopped being theoretical

The tooling pressure is not just anecdotal. In the arXiv paper How AI Coding Agents Modify Code, researchers analyzed 24,014 merged Agentic PRs with 440,295 commits and compared them with 5,081 merged Human PRs with 23,242 commits using the MSR 2026 Mining Challenge version of the AIDev dataset. They found substantial differences in commit count, reporting Cliff’s δ=0.5429, plus moderate differences in files touched and deleted lines. The study also found that Agentic PRs had slightly higher description-to-diff similarity across all measures.

Those are not morality stats. They do not say agent PRs are good, bad, or secretly three squirrels in a hoodie. They say agent contributions have measurable shape, and tooling should expose that shape rather than compressing it into a vibes-based green check. A good VS Code extension for agent-generated PRs should make commit count, files touched, diff summary, review state, and security findings legible at the moment a human is deciding whether to trust the thing.

Review is where the bottleneck hides

According to the arXiv paper 3100 Opinions on Code Review in an AI World, coding agents now author entire pull requests, and practitioners disagree about whether code review becomes the bottleneck, whether human review remains necessary, and whether review loses the understanding it used to build. The same paper says an observational analysis of public GitHub activity found agent-authored pull requests were reviewed less often, merged several times faster, and discussed less than human-authored ones. Then comes the fun little cactus in the chair: the direction of those trends flips under different but equally defensible analysis choices.

That is the dashboard trap. If your extension merely counts faster merges and fewer comments, it may reward exactly the behavior your team should inspect more carefully. The arXiv study collected 38,709 grey-literature documents, coded a stratified random sample of 3,100 with an LLM-assisted pipeline, and built a causal model with 26 constructs and 67 relationships. Translation for builders: instrument the workflow, but do not pretend the workflow is self-explanatory. Metrics are smoke alarms, not judges in tiny robes.

IDEs and terminals are converging on agent operations

The other side of this story is that agents are not politely staying inside the editor. In the arXiv paper Building AI Coding Agents for the Terminal, the authors present OpenDev as an open-source command-line coding agent and describe architecture choices including workload-specialized model routing, a dual-agent design separating planning from execution, lazy tool discovery, adaptive context compaction, automated memory, and event-driven system reminders. That is a lot of machinery for what used to be called typing commands, but here we are, narrating shell scripts like wildlife documentaries.

That matters for VS Code because editor tooling has to meet agents wherever they originate. Some work will begin in a terminal-native agent, some in an IDE extension, some in a PR bot, and some in whatever cursed integration your team wrote during a sprint named after a Greek letter. Tech Insider’s tutorial is useful because it treats the editor as an aggregation surface for agent work, not as the only place agent work happens. The winning pattern is not a prettier chat box. It is an operations view for code changes created by increasingly autonomous tools.

For builders, the immediate move is modest: surface agent PRs where review already happens, keep human review status visible, and preserve enough context that speed does not become archaeology. Watch whether editor release notes keep naming agent session scale, because release notes are where product truth goes after the keynote confetti has been vacuumed. The editor is no longer just where code is written. It is where the interns are counted.