Firebender

The deciding trait is that it looks at Compose previews while it works. An agent that can see the rendered @Preview it just changed is answering the question you actually have, which is whether the screen looks right, instead of asserting that the code compiles and leaving you to run it. Nothing else in the JetBrains column does this.

Outside Android that advantage disappears and you are paying for a competent general agent. Pick it if your day is mobile UI. Pick DevoxxGenie if it is not.

The dependency is the risk. Symbol lookup through a language server is more accurate than text search until the index is stale, the project is mid-migration or a Gradle sync has failed, and those are precisely the states an engineer reaches for help in. A tool that is strongest on a healthy repository is weakest on the day it is needed.

What it does right is isolate parallel attempts in worktrees, so two competing approaches cannot write over each other, and abandoning one costs a directory rather than a revert.

1. Most agents reason about a defect from the code that contains it, which is the same information the author had when they wrote it. Taking evidence from a running process changes the epistemic position entirely, because a stack trace and a variable value are observations rather than inferences.

2. Checkpointing supplies the other half, since a loop that acts on runtime evidence needs a way back when the evidence was misread. 3. No measurement of either is published, so the design is persuasive and unverified, which is the normal condition of this catalogue.

A ladder that wide tells you the company has learned that its heaviest users are worth sixteen times its lightest, and it has the confidence to charge for it. That is pricing power, and it is earned by being the only serious option inside one ecosystem rather than by being the best agent overall.

Moat: the mobile toolchain, which is defensible while a single vendor owns that IDE. Likely acquirer: whoever wants a mobile developer footprint without building one. Position: good product, narrow base, and the base is the risk.

The arithmetic is clean and that is worth something: a per-user rate with included usage means I forecast a fixed figure rather than a meter, and finance prefers a fixed figure. What it does not tell me is whether a directory integration, an audit export or a retention policy exists, because none of that is on the pages I can read.

There is no unattended mode either, so this improves how engineers work and never becomes a build step I can measure. Approved with conditions: a security questionnaire answered in writing before the contract.

Charging me before I may use my own credentials is the part I cannot talk myself past. The cheapest plan unlocks it, which is honest enough, but the principle stands: a key I pay for should not need a licence to work. Nothing here runs against weights on my own hardware either.

The redemption is the routing layer. Pointing it at a LiteLLM deployment means I decide what sits behind the endpoint, including things the vendor never listed, and my own servers attach as tools. That is more room than a closed product usually leaves.