AI Coding Agents Move From Chat Windows Into Infrastructure
AI coding agents are quickly shifting from novelty chat interfaces into core parts of developer workflows. Companies like Cursor and Amp see agents not just as helpers inside an editor, but as programmable infrastructure that can run tests, manage code changes, and orchestrate long-running tasks. This vision is driving a new wave of developer tools evolution, where interfaces become as important as models. At stake is a fundamental question: will the primary entry point for AI coding agents be the traditional terminal, cloud dashboards, or integrated development environments? Cursor is extending its AI-powered editor with a dedicated SDK that lets teams build and host agents on its runtime. Amp, meanwhile, is rebuilding its CLI, Neo, to act as a remote-controllable hub for agents that live mostly in the cloud. Their differing approaches expose both the promise and friction of an emerging CLI agent interface era.
Cursor SDK: Powerful Harness, Early-Stage Constraints
Cursor’s new SDK packages the same runtime, harness, and models that power its AI editor into a toolkit for building agents. The harness handles chores such as connecting to MCP servers, managing skills, observing and extending the agent loop, and coordinating subagents for delegated tasks. For teams wanting to run many agents in parallel from both the editor and the CLI, this offers a way to avoid managing virtual machines or hitting local memory limits. Yet the Cursor SDK limitations are significant. The public beta only supports TypeScript, with no official Python support; Python users must instead call a separate cloud REST API. Tool-call schemas and APIs are still in flux, and even advocates advise using it first for low-risk automations. Cursor’s design clearly targets agents embedded directly into existing workflows like CI, internal tools, and code review, but it also underscores how unsettled the agent ecosystem remains.
Amp Neo CLI: Turning the Terminal Into a Cloud Control Surface
Amp’s Neo CLI takes almost the opposite tack: instead of pulling agents into the editor first, it reimagines the terminal as one of several control surfaces for cloud-based agents. Neo’s core change is architectural. The agent loop now runs in the cloud, while the CLI acts as a thin client that streams updates and sends commands. Developers can start a local thread and then monitor or control it remotely via Amp’s web interface, queuing prompts, interrupting tasks, or cancelling agents from outside the command line. Neo adds a plugin system, better handling of long-running sessions through a compaction-first design, and visibility into intermediate reasoning, token usage, and cost. Amp argues that while agents are moving beyond a single terminal session, the CLI agent interface still matters for moments when developers want the agent “right next to” them during hands-on work.
Terminal-Centric vs Cloud-First: Competing Visions for Agent Workflows
Cursor and Amp embody a broader industry tension: should AI coding agents primarily live where developers already type commands, or in persistent cloud environments orchestrated through multiple touchpoints? Cursor pushes agents into familiar workflows—CI pipelines, maintenance scripts, code review—through a runtime tied closely to its editor and CLI, but still constrained by early-stage APIs and TypeScript-only support. Amp’s Neo instead prioritizes remote control and long-lived sessions, treating the terminal as a window into cloud agents rather than their home. Other tools, from GitHub Copilot CLI to Claude Code, are also adding remote capabilities, suggesting convergence toward hybrid models. Yet standardization is still missing: schemas are unstable, best practices for permissions and secret handling are evolving, and teams must decide where human review fits. The result is a fragmented, fast-moving landscape where both terminal-centric and cloud-first paradigms are actively competing for developer mindshare.