What does not work yet

Every agent platform demo looks the same: someone types a sentence, something impressive happens, everyone claps. None of them tell you where the thing falls over. This post is that list for Kavela, written while it is still short enough to be embarrassing.

We are early. Bootstrapped, small, shipping fast. If you are evaluating us against something more mature, some of the below will decide it for you, and it is better you learn it here than three hours in.

Things that genuinely work

Worth stating plainly so the rest has context.

You describe an agent in conversation and it gets built: skills matched to the task, connections, custom tools, and a system prompt that actually calls those tools. Runs execute server-side and survive you closing the tab. Each user of a published agent can connect their own accounts, and that separation is enforced at the database rather than by convention. Agents can run code and drive a real browser. You can publish an agent and get paid when people use it.

Things that are rough

The builder occasionally needs a second attempt. Our bar is that a build task completes conversationally without retries or dead-ends. We do not always hit it. When we miss, the usual cause is the builder not knowing how to wield a capability rather than the capability being broken, which means the fix is teaching rather than patching. That distinction matters to us and is invisible to you, so from the outside it just looks like a retry.

Cold starts are real. The first thing an agent does in a fresh session is slower than the tenth. Sandboxes hydrate, context loads. We have measured it, improved it, and it is still noticeable.

Long-running work is better than it looks but harder to observe. Runs survive disconnection, which is the hard part. Showing you clearly what a detached run is doing right now is the part we have not finished.

Things we have not built

No compliance certifications. We hold none. If your procurement process requires one, we cannot satisfy it today, and saying otherwise would be a lie you would discover later.

No local or on-premise version. It has been researched, not shipped. If your data cannot leave your infrastructure, we are not currently an option.

We are not a coding agent. If you want something to write software in your editor, other tools do that well and we are not competing for it. Kavela builds an agent that does a job for you or your users.

Marketplace economics are early. The mechanism works end to end: publish, price, subscribe, get paid. But there is not yet a large audience buying, so "publish an agent and earn" is a real capability and not yet a reliable income.

Why publish this

Two reasons, one principled and one selfish.

The principled one: an agent that acts on your accounts with your credentials requires trust, and trust starts with whether we tell you the unflattering things. A vendor who will not name a single limitation is telling you something either way.

The selfish one: the people we most want using Kavela are builders who will hit these edges in week one. Finding them here rather than mid-project is better for everyone, including us.

This list will change. When something on it gets fixed, the post gets updated rather than quietly deleted.