Cursor Cloud Agents: What to Give the Agent and What to Leave to the Platform (Category AI4SDLC)
I read Josh Ma’s June retrospective from Cursor on a year of building cloud agents. What caught my attention was not increased autonomy, but a shift in the architectural boundary: procedural logic is moving out of the agent harness and into tools controlled by the agent. Complexity does not disappear; it accumulates around the environment, reliability, policies, and state. A cloud agent is no longer a loop in one VM, but a system of workflows, environments, event logs, tools, and subagents.
1️⃣ The first shift: the environment became part of agent quality. Locally, the agent inherits repositories, dependencies, the build, tests, settings, and access from the developer’s laptop. In the cloud, all of this must be reconstructed explicitly. Cursor observed that an incomplete environment may not fail with an error: the agent simply performs worse, and the degradation is blamed on the model.
I would separate two roles. A sandbox limits actions, while the agent’s development environment contains everything required for the “change → run → verify” loop. An isolated VM without dependencies, tests, APIs, or a managed network may be safe, but useless.
Cursor therefore builds a resumable workspace: a VM can be suspended and resumed, images can be saved, restored, and branched, while the network and credentials are controlled separately. In a related article, the company describes the environment as code: a Dockerfile, version history and rollback, auditing, egress rules, and rules for handling secrets. This is already platform engineering for agents.
2️⃣ The second shift: the harness stops dictating the route. It used to recheck the result, force a commit and push, and, in CI Autofix, collect logs as well. Now the agent receives GitHub CLI, tools for branches and PRs, a map of repositories, and searchable files containing large outputs—and chooses the procedure itself 🤖.
The deterministic layer remains. I would draw the boundary this way: the agent chooses the tools, their order, and when to verify; the platform enforces permission and network boundaries, secret issuance, recovery, auditing, and duplicate protection when external actions are retried. The harness stops being the scriptwriter: it provides capabilities, while guarantees are distributed between execution orchestration and environment management.
3️⃣ The third shift: “one agent” is decomposed by lifetime and ownership of state. The agent loop lives in Temporal, the VM lifecycle is managed separately, and conversation storage and streaming are moved into their own layer. When a step is retried, the client rewinds the displayed stream; an asynchronous subagent can work on another pod and outlive its parent.
The workflow itself was also split up: instead of one “eternal” process, there are several short ones, each ending with a single task; individual operations receive their own timeouts and retries.
For interface work (computer use), Cursor still has a separate subagent with its own model routing, instructions, and screen recording. VNC and Chrome live in the shared environment, while the parent decides when to involve it. A mature capability can be given to the agent as a tool; a weak one is still supported by additional harness logic.
In practical terms, this leads to three decisions for a platform team:
- Version the environment description: image, dependencies, readiness checks, network rules, and rules for issuing secrets; store and rotate the secrets themselves separately;
- Provide capabilities through tools, while keeping security, reliability, and auditing outside the model’s probabilistic plan;
- Separate the agent loop, machine, conversation, and subagents by state owner, lifetime, time limit, and restart rules.
My main takeaway is this: moving logic from the harness toward the model is not simplification, but a change of contract. The boundary must be moved separately for each capability, and only after quality checks (evals).
The smarter the agent, the less the platform dictates its steps—and the better it manages the environment, boundaries, and consequences.
#AI #AI4SDLC #Agents #Architecture #PlatformEngineering #Engineering