Cursor Agent SDK
Findings that cite notes
Cursor released a kit for building AI agents, and its example, a code reviewer that cites the note behind each finding, is known only from Cursor's account.
Agents · Checked 13 Nov 2026 · 14 sources · 9 min

Most workplaces keep a record of the times something went wrong. Someone restarted a pump before bleeding the air out of the line, and afterwards somebody wrote down what happened and what to do differently. The note went into a binder. A year later a new hire who never heard the story is about to do the same thing, and the note is still in the binder, unread. Writing a lesson down is the easy half. Getting it in front of the right person as they are about to repeat the mistake is the half that fails.
Software teams have the same trouble, and one natural place to catch it. A programmer who wants to change a shared program proposes the change, and the proposal lists exactly which lines would be added and which removed. That listing is called a diff, and the proposal as a whole is called a pull request, or PR. A colleague reads the diff before the change is accepted. That reading is only as good as the reviewer's memory, and a reviewer who joined after the accident can let the same mistake through.
Cursor, a company that sells AI software to programmers, released a product called the Agent SDK in a post by Rustam Lalkaka dated 20 October 2026. SDK is short for software development kit, a set of ready-made parts that programmers build with. The post's worked example is a reviewer called kb-reviewer, built by one of Cursor's own teams. By Cursor's account it runs on every pull request in that team's main body of code, flags changes that contradict a recorded lesson, and cites the note behind each finding.
All of that is Cursor describing Cursor. A company introducing its own product picks the example and decides what to say about it, and the bad days are not the ones that get written up. What settles such an account is a stranger with nothing to sell running the same thing and reporting back. No independent review or benchmark of the SDK was found, a benchmark being a standard scored test. One outside page, from Enterprise DNA, described kb-reviewer as Cursor's own example, not yet evidence of results for other businesses.
A note the model never saw
The part of such a reviewer that does the reading is a language model, the kind of program behind chat assistants. A model does one thing: handed some text, it writes text in reply. It knows what it took in when it was built, which did not include one team's private notes. It cannot open a folder to look for them or post a comment on a pull request. Left to itself, the most it can do with a diff is write some remarks that go nowhere.
So a second program is wrapped around the model, an ordinary one that follows fixed steps. Suppose a team's notes include this lesson: never retry a card payment automatically, because a customer was once charged twice. Suppose too that a newcomer opens a pull request that makes the payment step try again whenever the bank is slow to answer. The wrapper hands the model three things: the team's standing instructions, which say to check every change against the lessons; a list of actions the model may ask for; and the diff.
The model can't search the notes, but it can write a request for a search, in a fixed format the wrapper understands: look in the notes for anything about retrying payments. The wrapper runs the search, which is plain code, and passes the results back to the model as more text. Among them is the double-charge note. Now the model has the change and the lesson side by side, and it writes its finding: this change retries payments automatically, which the note on double charges says never to do. The wrapper posts that as a comment on the pull request.
Each action the model may ask for is called a tool, and the model's part of the work, reading what it is handed and writing a reply, is called a turn. The cycle in which the model asks for a tool and gets its result, repeated until it writes an answer in place of a request, is the model loop, and the wrapper that runs it is called a harness. A model set up this way, with instructions, tools and a harness, is commonly called an agent. In the payment case the model could name its note because a tool had fetched the note and put it in front of the model. A model can also write a confident sentence that is wrong, and a finding that names its note can be tested by opening the note.
The payment case is made up, to show how such a reviewer could work. Whether kb-reviewer works that way is not known: the post does not say how it matches a change to a lesson or picks the note it cites, and its code, its instructions and its knowledge base are unpublished. None of the Agent SDK documentation pages linked here names a built-in citation feature. The nearest thing they describe is a set of review templates, each "one model turn" that "reads the diff and posts one comment".
The loop is the same whatever an agent is for. What changes from one agent to the next is the instructions and the tools, and Cursor's kit is built on that split. In its documentation an agent is a folder of files, and where a file sits decides what it is: a file named instructions.md holds the standing instructions, and each file in a folder named tools is one tool. The developer writes the instructions and the tools, and Cursor runs the model loop. The documentation overview says the SDK "uses the Cursor SDK harness that powers Bugbot, Security Review, Approval Bot, and Rollouts", Cursor's own products.
Every pull request, no count
By that division, building a reviewer means writing its instructions and its tools, and the post says the Cursor team built kb-reviewer "in about a day". It gives no team size and no starting point. Once built, an agent is started by an event, and Cursor's changelog entry of the same date lists pull requests among the events that can. The post says kb-reviewer runs on "every PR in their main codebase".
The post stops at "every". It gives no count of reviews, of findings, of findings the team accepted, or of false positives, meaning findings against a change that breaks no lesson. In the payment case, a false positive would be the reviewer citing the double-charge note against a change that only rewords a label on the payment page. A reviewer that does that often can teach people to skip its comments, and a true finding can be skipped with the rest.
Counting right from wrong takes its own kind of test. A model given the same diff twice may word its answer differently each time, so a test cannot demand an exact match the way a test of ordinary software can. The SDK's unit of testing, called an eval, "sends a fixed input to the real agent" and checks which tools ran and what the answer looks like. In the payment case that could mean handing the agent the retry change, then checking that the note search ran and that the answer names the double-charge note. One kind of check is a model judge, a model that scores the answer against a minimum. The announcement gives no benchmark, accuracy figure, speed claim or price.
A wrong finding costs someone's time. A wrong action can cost more, and tools are where a model's text becomes action. One setting, local.sandbox, is off by default. With it off, the reference says, commands the agent runs on its host computer are approved automatically and inherit the host's credentials, the passwords and keys kept on that machine. A reviewer raises the stakes, since a model takes its direction from text and a diff is someone else's text. The documentation advises turning the setting on when turns read untrusted input such as PR diffs.
Open source, all rights reserved
The documentation installs the kit from npm, a public registry of code packages, under the name @grok/agent-sdk. The name in the post's own code sample, @cursor/agent-sdk, returned a not-found error there on 13 November 2026, and neither the post nor the documentation overview explains the change of name. The package's README, the page that introduces it, calls the SDK "the open-source version of the engine behind" the same Cursor products. Open source has a settled meaning: the code is published under a license that lets anyone read it, change it and pass it on. The license file in version 0.6.0 says "All rights reserved. Use is subject to Cursor's Terms of Service." That is the opposite arrangement, called proprietary, in which the owner keeps those rights and sets the terms. No public repository, the kind of open store where such code is normally kept, is confirmed to hold the SDK's source code.
Where an agent runs is a separate matter from who owns the kit. The post describes sending a finished agent to run on Cursor's own computers, which the command reference says needs a team with agent hosting turned on. The changelog entry adds that an agent can run on the developer's own infrastructure, called self-hosting. But the configuration reference defines the model setting as the "Cursor model for turns", and the documented self-hosting example sets a Cursor credential, the secret that identifies an account with Cursor. Only two commands, run and eval, need that credential, but as documented, a self-hosted agent still sends its model turns to Cursor. In the payment case the agent would run on the team's own computer, and the model's turns would still travel to Cursor and back.
On Cursor's account, teams can write pull request reviewers and approvers in the form Cursor uses for its own bots, as files that pass through review like any other change. On the documented terms, what they build depends on Cursor for model turns and on Cursor's Terms of Service, whether it is hosted or self-hosted. One commenter on the forum Hacker News named being tied to a single vendor as a reason to want a vendor-neutral standard. The outside coverage found is forum comments and small-site rewrites, none from established trade press, and no cited source reports hands-on use of the SDK.
Check it yourself
- Open Cursor's announcement and read the kb-reviewer paragraph. It confirms the "in about a day" and "every PR in their main codebase" wording, and that no count of findings comes with them.
- Open the npm record for @grok/agent-sdk and check that the package exists under that name. The same path with @cursor/agent-sdk returned not found on 13 November 2026.
- Open the license file in version 0.6.0 and set it beside the README's "open-source version". It confirms the all-rights-reserved terms.
- Open the review templates page and read how each template is described. It confirms the "one model turn" that "reads the diff and posts one comment".
- Read the model and local.sandbox entries in the configuration reference. They confirm the "Cursor model for turns" wording and the sandbox setting being off by default.
- Read the evals guide for what an eval checks. It confirms the fixed input sent to the real agent, the check on which tools ran and the model judge.
- Work through the Agent SDK quickstart with a Cursor credential. No cited source reports hands-on use, so this is the only check of how the SDK behaves.
Sources
- Introducing the Agent SDK. Cursor, dated 20 October 2026.
- AI Pulse report on Cursor's Agent SDK. Enterprise DNA, 28 October 2026.
- Review and triage templates. Cursor.
- Agent SDK project layout reference. Cursor.
- Agent SDK, documentation overview. Cursor.
- Agent SDK, changelog entry. Cursor, dated 20 October 2026.
- Agent SDK evals guide. Cursor.
- Agent SDK agent config reference. Cursor.
- @grok/agent-sdk, package record. npm registry.
- LICENSE.md in @grok/agent-sdk 0.6.0. unpkg.
- Agent SDK CLI reference. Cursor.
- Build with a coding agent. Cursor.
- Hacker News comment, item 50274022. Hacker News API.
- Agent SDK quickstart. Cursor.