A useful internal agent eventually needs the same discipline as other production software: a versioned definition, repeatable tests and a way to inspect what changed. A prompt saved in a chat is a poor substitute for that operational record.
Cursor’s October 20 Agent SDK release describes a project containing instructions, typed tools, skills, MCP connections and message channels. The SDK is generally available, supports local development and evaluation, and can deploy to Cursor or the customer’s infrastructure. It also records conversations and tool calls within sessions.
Test the action contract
The release describes evaluations that assert on called tools, their inputs and replies. Those checks can catch an agent using the wrong operation or omitting a required step. They do not, by themselves, prove a downstream system reached the desired state. A test that sees a deployment call should still inspect the deployed result.
The ability to replay GitHub events is useful for examining a workflow against a reproducible trigger. A convincing test set also needs adverse cases: repeated events, missing permissions, stale repository state and failed external services. Otherwise an automated response can look reliable because the evaluation only contains its easiest path.
A repository makes review possible
Packaging behavior as code gives reviewers a concrete place to inspect tool authority and integration changes. It also introduces normal maintenance obligations. Secrets, runtime permissions and the installation’s event subscriptions must match the reviewed project rather than drifting independently.
The immediate benefit is a more inspectable delivery process for specialized agents. Broader gains depend on whether teams evaluate accepted outcomes and keep deployment configuration aligned with source. The next evidence to watch is whether a changed agent version passes realistic failure tests and produces the correct external artifact, with less review effort than the existing workflow.
Sources & evidence
Sources describe the announcements and evidence available when this analysis was published. Company claims are attributed; plans and trials are not treated as sustained operation.
