From IDE Add-ons to Shared AI Coding Agent Runtimes
AI developer tools are rapidly evolving from editor-specific assistants into shared infrastructure powered by standardized agent runtimes. Instead of tightly coupling AI coding agents to a single IDE or one-off plugin, emerging platforms are exposing their internal “agent loops” as reusable SDKs that any team can build on. This shift turns an AI coding agent runtime into a first-class piece of a team’s stack, rather than a black box tucked inside a proprietary interface. It also lowers the barrier for organizations to create proprietary AI coding agents aligned with their own workflows, compliance rules, and codebases. The trend spans both open-source and proprietary offerings, but the common thread is clear: runtimes and harnesses are becoming the primary product, while editors, CLIs, and web consoles act as interchangeable control surfaces layered on top.
Cline SDK: An Open-Source Agent Runtime for Every Surface
Cline’s @cline/sdk is an open-source agent SDK that now powers all of Cline’s own surfaces, from its CLI to Kanban and, soon, VS Code and JetBrains extensions. Rather than extending an IDE-bound architecture, Cline rebuilt its AI coding agent runtime as a layered TypeScript stack. Foundational types live in @cline/shared, model providers like Anthropic, OpenAI, Google, AWS Bedrock, Mistral, LiteLLM, and OpenAI-compatible endpoints are isolated in @cline/llms, while @cline/agents runs a stateless agentic loop that coordinates tools and emits events. On top, @cline/core manages session lifecycle, persistence, and configuration, allowing sessions to move across interfaces without being tied to a single UI. Teams can consume the full AI coding agent runtime via a single npm install or cherry-pick packages for a narrower surface, turning Cline into an open-source agent SDK rather than just an IDE plugin.
Cursor SDK: Cloud Runtime Power with Early-Stage Constraints
Cursor is pushing in a different direction with a proprietary AI coding agent runtime exposed through its SDK. The company’s harness—responsible for executing tests, managing tools, and evaluating performance—now backs both the Cursor editor and custom agents built by users. Developers gain access to infrastructure like MCP server connections, automated skills management, hooks into the agent loop, and subagent delegation, all on the same cloud runtime Cursor uses internally. That means teams can avoid managing their own VMs or worrying about memory ceilings while still running many agents in parallel via the editor or CLI. However, the Cursor SDK remains in public beta and is currently TypeScript-only, with Python developers advised to hit a Cloud Agents REST API instead. Cursor advocates urge teams to start with low-risk tasks, emphasizing that the real challenge is governing what agents may change and when humans must review their work.
Amp Neo: A Remote-Controllable CLI for Agentic Workflows
Amp’s Neo CLI reimagines the terminal as a control surface for long-running AI coding agents rather than their primary execution environment. The rebuilt CLI introduces remote control: developers can start an agent thread locally, then monitor and steer the same session from a web interface that streams terminal updates to the browser. From there, they can send follow-up prompts, queue messages, interrupt tasks, or cancel the agent entirely without living inside the command line. Architecturally, Amp moved the agent loop into the cloud, cutting data transfer while enabling more durable, cross-environment workflows. Neo also layers in a plugin system and a compaction-first approach to managing large conversational histories, signaling a shift from traditional shell utilities to an AI-native runtime. Even as Amp argues the “coding agent” in its old form is obsolete, it insists the terminal still matters as a familiar, flexible touchpoint for agent interaction.
Cloud vs Terminal—and Why Open Runtimes Tilt the Field
Across Cline, Cursor, and Amp, the debate is no longer about whether AI coding agents matter, but where they primarily live. Cursor leans into cloud-hosted runtimes accessed from IDEs and CLIs; Amp moves the loop to the cloud while treating the terminal as a remote control panel; Cline decouples its open-source agent runtime from any one surface so sessions can flow across CLI, boards, and IDEs. The emerging consensus is that interfaces—cloud dashboards, terminals, editors—are interchangeable, while the AI coding agent runtime itself becomes the strategic asset. Open-source agent SDKs like @cline/sdk reduce lock-in and let teams standardize their own tools, while proprietary runtimes like Cursor’s and Neo’s push capabilities that may be hard to replicate quickly in-house. Together, these platforms are normalizing shared agent harnesses, making it far easier for engineering teams to design, own, and evolve bespoke AI coding agents.
