Product / training feedback — not a bug report.
One-liner
Instrument and train on moments where a user successfully corrects the agent
(wrong path, missed metaphor, overbuilt approach, false assumption) — anonymized —
as a first-class learning signal. These events are much denser than generic chat
tokens and could materially speed short-horizon RSI / product quality.
The problem
Today, when a user fixes the model’s map mid-session (“one level up” meant ../repo,
not a filesystem scavenger hunt; “1 fps” meant frame-at-a-time not sample rate = 1),
the value is mostly wasted:
- the agent may adapt for the rest of that chat
- the base model / product loop often gets nothing structured back
- the same class of error repeats for the next thousand users
That’s leaving free, high-precision labels on the table.
Proposal
Detect / let users mark “successful correction” events and ship anonymized
training/eval records, e.g.:
-
Trigger (any of):
- explicit user action: “This corrected you” / thumbs on the correction turn
- heuristic: user points out a misunderstanding → agent acknowledges → subsequent
behavior changes in-session - optional: agent self-tags “user successfully refined my interpretation of X”
-
Payload (anonymized, redacted, opt-in or privacy-mode aware):
- error class taxonomy (pathing, metaphor/analogy, scope, tool overuse, API assumption…)
- short before/after intent paraphrase (no private code/secrets)
- environment tags only (OS, agent mode) — not identity
- optional embedding of the misunderstanding type for clustering
-
Downstream use:
- preference / DPO-style pairs (wrong interpretation → corrected interpretation)
- eval harness: “does the model make this mistake class less often?”
- product routing: if “spatial path clue” cluster spikes, fix prompts/tools
Why this is high leverage (reCAPTCHA analogy)
reCAPTCHA turned tiny scraps of human recognition into infrastructure-scale signal.
Agent corrections are similar: 10 seconds of a sharp user’s brain often encodes a
hard world-model update (geometry of folders, meaning of a joke, what “go” meant).
You don’t need more raw tokens — you need more of these events, labeled and fed back.
A lightweight “correction log → anonymized feedback pipeline” could 10–100× the
usefulness of existing thumbs/feedback for comprehension bugs specifically.
Concrete UX (minimal)
- In chat: “Mark as correction” on a user turn (or on the agent’s acknowledgment)
- Weekly opt-in: “Help improve Cursor with anonymized correction events”
- Dashboard for the lab: top misunderstanding clusters this week
What I’m not asking for
- Not “the model rewrites its weights in my chat”
- Not storing private repos by default
- Not replacing human eval — amplifying it
Motivation
Heavy Cursor user. Watching agents fail geometrically (path, metaphor, over-search)
and get corrected cleanly feels like the highest-SNR training data in the product —
currently under-instrumented compared to “was this code answer good?”
Happy to clarify taxonomy or examples if useful.