How Cursor sets up their cloud agent environment

6 hours ago

The development environment is a product for agents, and must be designed to match cloud and local development.
To enable cloud agents, Cursor made the repo work on Ubuntu VMs, added dependencies to a Dockerfile, and implemented security features for secret management.
A custom CLI called 'anydev' simplifies build and run commands, making it easier for agents to test code reliably.
Cursor Cloud MCP was built to allow agents to inspect and heal their own environments, detecting setup failures and other issues.
An automation called 'Cloud Doctor' periodically checks for failures, does root cause analysis, and opens PRs to fix issues.
Cloud Doctor also analyzes agent traces to identify misleading skills or slow workflows, then improves the environment for future agents.
As a result, cloud agents now author more than half of Cursor's merged PRs, up from one in ten in December.
To assess codebase readiness for cloud agents, consider whether agents have access to the same tools and data as developers, can find relevant skills, and can test core workflows.

Cloud agents evolved from simple extensions of local agents to independent VMs with their own environments, enabling parallel, unattended, and long-running tasks.
A full development environment is critical for cloud agent output quality, requiring infrastructure like user tools, VM hibernation, and pipelines for checkpointing and restoring images.
Cloud agents face reliability challenges due to disruptions like inference outages; using Temporal for durable execution improved reliability and now handles millions of actions daily.

Conductor Cloud launched on July 30th, shifting coding-agent management from local repositories to persistent cloud workspaces.
Developers can collaborate in shared workspaces, prompt agents together, and monitor work via a desktop client; the API enables programmatic workspace and agent management.
Each cloud workspace runs in an isolated microVM with preinstalled dependencies, providing a persistent environment for long-running tasks.

Knock Agent is an AI agent for managing customer messaging resources, capable of creating and updating workflows, templates, audiences, segmenting users, and answering performance questions.
The agent was built using a virtual filesystem and bash, inspired by Vercel's approach, allowing it to explore and script on resources via a filesystem instead of exposing numerous API tools.
A custom Elixir-based virtual bash interpreter called just-bash was developed to provide a lightweight, in-memory environment without the performance overhead of a full sandbox.

The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.

Sentinel is an open-source QA agent that reads the codebase to understand business flows and performs end-to-end testing across frontend and backend.
It autonomously derives critical business flows from the code, such as reservation lifecycle, group bookings, and payment processes, similar to a human QA engineer.
During testing, Sentinel interacts with both UI and API, catching backend bugs that UI-only tests miss, like state-machine inconsistencies and API/UI disagreements.

Cloud agents evolved from simple extensions of local agents to independent VMs with their own environments, enabling parallel, unattended, and long-running tasks.
A full development environment is critical for cloud agent output quality, requiring infrastructure like user tools, VM hibernation, and pipelines for checkpointing and restoring images.
Cloud agents face reliability challenges due to disruptions like inference outages; using Temporal for durable execution improved reliability and now handles millions of actions daily.

Conductor Cloud launched on July 30th, shifting coding-agent management from local repositories to persistent cloud workspaces.
Developers can collaborate in shared workspaces, prompt agents together, and monitor work via a desktop client; the API enables programmatic workspace and agent management.
Each cloud workspace runs in an isolated microVM with preinstalled dependencies, providing a persistent environment for long-running tasks.

Knock Agent is an AI agent for managing customer messaging resources, capable of creating and updating workflows, templates, audiences, segmenting users, and answering performance questions.
The agent was built using a virtual filesystem and bash, inspired by Vercel's approach, allowing it to explore and script on resources via a filesystem instead of exposing numerous API tools.
A custom Elixir-based virtual bash interpreter called just-bash was developed to provide a lightweight, in-memory environment without the performance overhead of a full sandbox.

The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.

Sentinel is an open-source QA agent that reads the codebase to understand business flows and performs end-to-end testing across frontend and backend.
It autonomously derives critical business flows from the code, such as reservation lifecycle, group bookings, and payment processes, similar to a human QA engineer.
During testing, Sentinel interacts with both UI and API, catching backend bugs that UI-only tests miss, like state-machine inconsistencies and API/UI disagreements.

Hasty Briefsbeta