AI Agents Move From Novelty to Infrastructure
AI agent development is shifting from experimental chatbots to tools that sit deep inside the software delivery pipeline. Companies like Cursor and Amp are positioning agents not just as coding assistants, but as programmable infrastructure that can run tests, refactor code, and manage routine maintenance. Cursor’s leadership frames this as a “third era” of software development, where agents operate alongside continuous integration, internal tools, and automation scripts. At the same time, Amp argues that the familiar “coding agent” tied to a single editor or terminal session is starting to break down, replaced by more autonomous systems that run across environments and over longer time spans. Yet as these platforms race ahead, they expose unresolved questions: how to integrate agents safely into production workflows, which interfaces will dominate, and whether today’s tools are stable enough for teams to bet core processes on them.
Cursor SDK: Powerful Harness, Persistent Python Gaps
Cursor’s new SDK gives developers access to the same harness and cloud runtime that power its AI code editor, promising to remove much of the “agent stack” overhead. It automates MCP server connections, manages agent skills, exposes hooks into the agent loop, and coordinates subagents for specialized tasks. Early adopters highlight its potential for running many agents in parallel across editors and CLI tools without managing virtual machines or memory limits. However, important limitations hold back broader adoption. The SDK is TypeScript-only in its public beta, leaving Python users to integrate via a separate Cloud Agents REST API. Tool call schemas and APIs are described as unstable, and developers are urged to start with low-risk tasks such as fixing tests or summarizing changes. The result is a compelling but still-moving platform that underscores both the promise and the fragility of current AI agent tooling.
Amp’s Neo CLI and the Rise of Remote-Controlled Agents
Amp’s redesigned Neo CLI reflects a different response to the same pressures shaping AI agent development. While the company insists that the traditional coding agent model is fading, it also declares that “the terminal still matters,” reframing it as a control surface rather than the primary execution environment. Neo is built for remote control: developers can start an agent thread locally, then manage it via a web interface that streams live updates and accepts follow-up prompts or interrupts. Architecturally, Amp moves the agent loop into the cloud, reducing data transfer and enabling longer-running, more resilient workflows. A new plugin system and “compaction-first” design target the realities of extensive conversational histories and complex toolchains. Neo’s ability to expose intermediate reasoning and track costs inside the interface also speaks to a growing demand for transparency and control in CLI tools for AI, especially as sessions stretch beyond a single terminal window.
Terminals vs. Cloud: Competing Control Surfaces for Agents
The divergence between Cursor and Amp illuminates a broader debate: should AI agents be driven primarily from local terminals, code editors, or cloud dashboards? Amp’s Neo suggests a hybrid world, where the terminal remains vital but acts as one of several entry points into a cloud-hosted agent loop. Cursor, meanwhile, brings agents closer to where developers already work—CI pipelines, internal tools, GitHub issues—without prescribing a single dominant interface. Other tools like GitHub Copilot CLI and Claude Code are likewise experimenting with remote-control capabilities, reinforcing the idea that agentic workflows must span contexts. For developers, this raises practical questions about observability, collaboration, and security. If agents can be steered from anywhere, teams must decide where authoritative control lives, how interventions are audited, and which environment—CLI or browser—becomes the default cockpit for debugging and directing autonomous behavior.
Innovation Outpaces Stability in the AI Agent Toolchain
Both Cursor and Amp highlight the growing tension between rapid innovation and production-ready stability in AI tooling. Cursor’s SDK shows how far the ecosystem has come: agents that understand test suites, coordinate subagents, and plug into existing workflows. Yet its public beta status, unstable schemas, and lack of first-class Python support reveal gaps that make many teams hesitant to trust it with critical code changes. Amp’s Neo, rebuilt from the ground up to support remote control and plugins, demonstrates a similar pattern: ambitious architectural bets launched while the broader question of the “right” interface remains unsettled. For now, practitioners are advised to treat these platforms as powerful but evolving tools, best suited for low-risk automation and experimentation. Until language support gaps narrow and APIs stabilize, AI agent frameworks will remain on the edge of production, promising transformative gains but demanding cautious integration.