Cursor

Facts

Openness

Closed, studied from outside

Licence

proprietary

Language

Not stated

Pinned version

Official documentation (the code is closed)

Topics checked

31 of 33

Kind of agent

Coding

Project page (opens another site)

What it does notably well

• Compaction

  Cursor's 2026-03-17 post describes Composer self-summarization trained with reinforcement learning: it summarizes its own context at a token trigger, and summaries are rewarded…

• Human approval

  The Auto-review classifier runs in the parent's RPC stream and, on a block, sends an explanation to the parent agent, which can choose a safer path before any user prompt (vendor…

• Crash recovery

  Cursor's 2026-06-02 post says it decoupled the agent loop, the machine and the conversation state, so the loop lives in Temporal rather than on the VM.

• Result clearing

  Cursor's 2026-01-06 post describes dynamic context discovery: long shell and MCP outputs are written to files the agent reads on demand instead of being truncated, chat history…

• Sandboxing

  Cursor's 2026-02-18 post says it changed Shell tool results to surface the sandbox constraint behind a failure, so agents recover from sandbox failures more gracefully.

• Stopping

  A stop hook can answer the end of a run with a follow-up message that Cursor submits as the next user message, giving loop-until-done flows with a per-script loop limit.