Amp’s Rush 2.0 promises faster, cheaper fixes

Amp has just rolled out Rush 2.0, repositioning its agent mode for small, focused code edits instead of broad codebase work. It swaps in GPT-5.5 (no reasoning), streamlines tools, and nearly halves runtime, with a modest cost bump.

Amp’s Rush 2.0 promises faster, cheaper fixes

TL;DR

  • rush 2.0 repositioning: Faster, lower-cost mode for small coding tasks; no longer general-purpose codebase agent
  • Model change: Uses GPT-5.5 (no reasoning), replacing Haiku 4.5; aims for minimal correct edit, focused check, stop
  • Best-fit tasks: Fix failing tests, match component styling, rename symbols across files
  • Avoid for: Transient bugs, architecture changes, migrations, complex features, unclear “done” criteria
  • Toolchain update: shell_command for search/read/verify + apply_patch for edits; removed grep/glob/create_file; subagents match setup
  • Evals + workflow: 44% solved (from 39%); $0.58/task (from $0.44); 1:32 runtime (from 2:59); deep solved 58%; pair with read-only oracle for review; run mode: use rush

Amp has released a new version of its rush agent mode, recasting it as a faster, lower-cost option for small coding tasks rather than a general-purpose agent for broader codebase work. The company states that rush now uses GPT-5.5 with no reasoning, replacing Haiku 4.5, and is meant to find the relevant files, make the smallest correct change, run a focused check, and stop.

Amp lists examples of the kinds of tasks rush is meant to handle: fixing a failing test, matching styling between components, and renaming a symbol across files. It also cautions against using the mode for transient bugs, architecture changes, migrations, complex features, or work where the definition of “done” is still unclear.

The release also changes the toolchain. rush now uses shell_command for searching, reading, and verification, plus apply_patch for edits. That allows Amp to remove several overlapping tools, including grep, glob, and create_file. The company adds that the task subagents use the same setup as the main agent: no reasoning and GPT-5.5.

On its internal evals, Amp reports that rush solved 44% of tasks, up from 39% for the previous version. Average cost rose to $0.58 per task from $0.44, an 18% increase, while runtime fell to 1 minute and 32 seconds from 2 minutes and 59 seconds. Amp characterizes that as a tradeoff of “a little more expensive, meaningfully more capable, and almost twice as fast.” It also notes that deep performed better on the same evals, solving 58% of tasks.

Amp also highlights a pairing with the oracle tool. The suggested pattern is to let rush build, then have the oracle plan, criticize, or review. The company describes the oracle as slower, more expensive, read-only, and suited to adversarial thinking. It suggests that approach for bounded tasks with enough complexity that a second pass may prevent a locally plausible but globally incorrect edit.

To use it, Amp says to run mode: use rush in the Amp CLI.

Source: Amp

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community