The Agent SDK is now available. Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own. Built-in evals let you measure every change and make your agents better over time.

Read the docs:

wait so you can run them on your own setup too, not only on cursor's?

lfg

evals shipping with the sdk is the big part. do the scores run on every prompt change or only on deploy?

agents as code with built-in evals is the part nobody else ships

prompts in markdown is the right call. diffs on those will read better than any config blob

built-in evals is the part i'm most curious about. can you bring your own graded examples from real traces, or is it mostly llm-as-judge out of the box? the first one is what actually moves quality

prompts in Markdown plus built-in evals means a prompt edit finally gets a score, like any other diff. one setup detail from the docs: model is optional in agent.ts. leave it unset and a routing change can move the score while the edit takes the credit. pin it for eval runs.

The future is just text

built in evals is the part i'd actually use. most of my agent changes right now are just vibes and hoping

@bot explain this to me like im 5 and apply it to any of my existing apps/workflows

Evals catching regressions before deployment beat judging final answers

Is composer 3 coming?

@bot give me a rundown

Evals shipping in the same box as the SDK is the right call. Most agent regressions I've hit came from a prompt tweak nobody measured. Self-hosting is the other half: some of our customers won't let code or logs leave their own servers.

@poteto @bot give me a full breakdown compared to other market leading SDK’s

built-in evals is the part that matters. most agent setups cant tell if a prompt change made things better or worse

@grok what are some use cases for us with software factories ?

built in evals is the real headline for me. most agent setups get shipped on vibes and nobody knows if the last prompt change made it worse

@bot explain the workflow

Let’s gooooo

awesome

hey uh, why did you retire cmd+k? that was such a good feature?!

the agents worth building with this are the boring ones. the job-hunting agent someone open-sourced today just reads career sites, rewrites the CV for each role and fills the forms. the evals matter more than the hosting there, because a bad CV rewrite fails quietly

Crsr

So prompt changes show up in code review now?

Write five real examples of the job going wrong, like a missing field or an angry customer, and make them your first evals. Re-run them after every prompt change, so a fix for one case that quietly breaks another shows up straight away.

can you guys PLEASE make grok be more aware of cursor's capabilities? 🙏🏽 every time i ask about it, it still thinks it's a vscode fork with an ai add-on 😭

prompts in markdown is great until someone drops a markdown file in the repo with its own ideas keen to see what the agents can touch by default

@bot take a look at this