History
How agent interfaces change
Documented timelines list only verified changes, each with a date and an official source. Concept timelines are illustrative teaching examples and are kept separate.
Documented
Verified · dated · sourced2025-12-10 · 1.107
Two-step approval for the fetch tool: approve the request (contacting the URL or domain), then separately approve using the fetched content. Trusted Domains skip the first step.
Our read — Outbound data risk and inbound prompt-injection risk get separate decisions.
code.visualstudio.com/updates/v1_1072026-10-15 · Docs (accessed)
Docs document request vs. response approval and per-pattern settings (approveRequest / approveResponse).
Our read — The split can be tuned per URL pattern instead of all or nothing.
code.visualstudio.com/docs/agents/run/approvals
Concept
Illustrative · not real product historyHow an approval prompt might mature, shown as a composite teaching sequence. These stages don't describe any specific product's releases.
Agent wants to edit 2 files
Stage 1
Yes / No
Without the target or diff, users can't judge risk, so they either approve everything or stop using the agent.
Agent wants to edit 2 files
src/components/Button.tsx
Stage 2
Show the diff
Decisions get better, but every edit still interrupts. Approval fatigue moves from 'blind' to 'tedious'.
Agent wants to edit 2 files
src/components/Button.tsx
Stage 3
Allow for session
Fewer interruptions, but the scope is invisible and the expiry is arbitrary. The same session can touch config and tests alike.
Agent wants to edit 2 files
src/components/Button.tsx
Stage 4
Scoped & visible
Interruptions now line up with real changes in risk, and users can always see what they've granted.