Amp 4.8 brings major improvements to agent control, code review, and model performance, with new capabilities for custom agents and remote execution.
Models and performance
- Opus 4.8 in smart mode: Achieves 62% task success (up from 52%), makes more focused changes, verifies work 15% more often, and proactively uses tools like librarian and repo skills.
- Opus 4.8 fast mode: Runs 2.5× faster at 3× lower cost than the previous version, preferring surgical edits over full file rewrites.
- Librarian performance boost: Now 3× faster and 43% cheaper on GPT-5.5, running 8 parallel tool calls per turn instead of 3, completing searches in about 5 turns rather than 15.
Agent control and customization
- Unified agent interface: Monitor and manage all active agents across web, mobile, and CLI in one place without juggling terminal tabs.
- Custom agents in plugins: Create custom agents in Amp plugins and use them as your main agent, subagents, or spawned worker agents you switch between, each with a unique orb color.
- Remote agent execution: Run agents in orbs, isolated machines that spin up per thread and bill by the minute, letting you launch multiple agents simultaneously without local resource constraints.
Code review and workflow
- In-app code review: Review code changes directly in the app on desktop and mobile with intelligent diffing that highlights only actual changes by detecting duplicate blocks.
- Interactive change staging: While a thread has an active environment, scroll through diffs, request changes on specific sections, and stage changes interactively.
- Asynchronous agent messaging: Threads let you send messages and wait for agent responses asynchronously, enabling workflows where spawned agents can respond back to parent threads.