Why self-assessment breaks in an agentic editor
Developers were already poor judges of their own productivity before agents arrived. In the one study that measured both subjective and objective productivity on the same population, the gap between them was more than 40 percentage points (METR / Becker et al., 2025; 16 developers, 246 tasks).
AI assistance widens the crack rather than closing it. When METR surveyed 349 technical workers in 2026, the median respondent reported roughly 3× on speed — and only 1.4–2× on the value of the work produced. Those are two very different numbers about the same working day, and the faster one is the one that gets remembered.
Speed is vivid. Value is not. Watching a diff appear in two seconds leaves a stronger impression than the forty minutes afterwards spent working out which parts of it you actually trust.
Those forty minutes are well documented. In Stack Overflow's 2025 developer survey of more than 49,000 respondents, the top frustration with AI tooling was code that is “almost right, but not quite” (66%), and 45% said debugging AI-written code takes longer than they expect. Trust in AI accuracy sat at 29–33%.
None of this is an argument against working this way. It is an argument that the feeling of a session has become a worse guide to it than it already was — and the fix is not a better feeling, it is a written-down record.
What NEDIO records, and what it refuses to
Before the block, in about ten seconds:
- the goal, in your words
- the work mode — code, debug, plan, review, or deep work
- how long you expect it to take
After it, in about thirty seconds:
- whether the goal landed
- how the block felt, in one tap
- one line on what actually shipped
- anything you are parking for later
What it never touches is the code. Not the file you had open, not the diff the agent produced, not the name of the repository. NEDIO cannot tell whether you typed a line or accepted it, and that is a design decision rather than a limitation: a tool that graded your acceptance rate would be measuring the agent, and you would start gaming it within a week.
Plan and review stop being overhead
In a conventional editor, plan and review are the modes people skip. They feel like the packaging around the real work.
In Cursor they frequently are the work. A block spent writing a precise specification and then reading back what came out is not a code block, and filing it as one quietly corrupts your own history — you end up comparing hours that have nothing in common and wondering why the numbers say nothing.
Tagging honestly costs one tap and buys the only thing that makes a weekly comparison meaningful: sessions that are actually comparable.
What you actually learn, and how long it takes
Not “Cursor made you 30% faster”. That question cannot be answered for one person — it needs a version of your week where you worked differently, and that week does not exist.
What a few weeks of sessions can support is narrower and more useful: that your plan-mode blocks land their goal far more often than your debug blocks, that your longest sessions finish worse than your middling ones, or that the same item has been on your parking line five weeks running.
And most weeks, honestly, nothing separable turns up at all. A single week of five sessions carries a 95% confidence interval of roughly ±43 percentage points on something as simple as a goal-landing rate — which is another way of saying one week is noise. NEDIO says so on those weeks rather than inventing a pattern to fill the screen.
Install
Search NEDIO in the Extensions panel, or open the Open VSX listing directly \u2014 that is the registry Cursor reads, and the build there is the same one the VS Code Marketplace serves. Then open the NEDIO icon in the activity bar and connect the editor once.
The panel, status bar, commands, and settings are identical to the VS Code build and are documented on the VS Code page.
Frequently asked questions
Is this a different extension from the VS Code one?
Same extension, different registry. Microsoft’s Marketplace terms cover VS Code itself, so Cursor reads Open VSX instead — every release is published to both, from the same build. The software is identical; this page exists because the reason to use it in Cursor is different.
Does NEDIO see what the agent generated?
No. NEDIO reads no workspace content at all — not your source, not diffs, not file names. It records what you said you were doing and what you said landed. Whether a human or an agent typed the characters is invisible to it, and deliberately so.
Will this tell me whether Cursor makes me faster?
Not on its own, and no honest tool would claim to. Answering that for one person requires comparing against a version of your week that did not happen. What NEDIO gives you is the raw material: what you intended, what landed, and how those two lined up across enough sessions to mean something.
What work mode should I use for agent-heavy sessions?
Usually plan or review. If most of a block was specifying work and reading back what came out, calling it "code" makes your own history harder to read later. The tag costs one tap and it is what lets AI Review compare like with like.
Does the music interfere with the Cursor chat panel?
No. Audio runs inside the extension and has no interaction with the editor UI. The sprint panel lives in its own activity bar container, so it does not compete for the sidebar slot you keep chat in.
Does it work in Windsurf, Trae, or other VS Code forks?
It is built and tested against VS Code and Cursor. Other forks that implement the VS Code extension API at version 1.85 or later will generally run it, but they are not something we verify before a release.