Kiro on X

The positive talk is short and unspecific: "Its good", "is the best", still using it for private projects, and one person who says a customer in production told them they run it. Several credit a faster model version at the same price ("a no-brainer upgrade") and the bug-fixing and code-review work it does on real projects. One posts that Amazon's front-end products, Kiro among them, are good but nobody talks about them.

5 people go further and show what it produced: an animation of real infrastructure captured from a live system, a data analysis across 285 college majors, a fraud detector for a dataset, an app carried over from another tool, and a three-hour hackathon demo that deployed.

The doubt is level with the engineering praise at 6 people each. The skeptics call the IDE "a nightmare to use" next to Cursor and Copilot, say it tests your patience, and one former fan blames the lag between a model's release and its arrival in the product. The engineering praise singles out recipes saved as reusable steps, a judge of plain code that decides which fixes ship, a sandbox that bounds what the agent can touch, and an agent that uses a company's own tools.

On price, people compare multipliers rather than quote a total: "at 2.4x multiplier" next to Opus 5.5 at 2x draws "bruh", "Damn 2.4? gesh. Why?" and "too pricey". One rules out using it over Opus 5.5 at that rate, and others guess at why one model costs more, including whether Anthropic is subsidising Bedrock or Kiro.

Billing trouble is the other half of the top. One user with several paid subscriptions through AWS reports usage at $205.65 and climbing despite overages being disabled; a reply explains that prices exclude tax and that subscribing in two regions bills twice. One person is out of credits until the end of the month, and AWS support answers someone whose subscription attempts failed.

Approvals are counted in clicks: "like 40 commands" per task, "always allow" that doesn't store, and even a simple find asking to be approved. Level with that, people report grep failing outside the workspace on Windows v3 and a CLI and IDE "full of ugly bugs", and 3 cannot get in — a suspended student account, an enterprise user who says it isn't available, and a request for access. 2 trade JetBrains install workarounds.

The comparison with other editors is unfavourable and specific. One person, with the most reach in the section, calls the IDE "a nightmare to use bro, even when compared to Cursor and Copilot", and objects to approving around 40 commands because "always allow" doesn't store the approval. The other weighs cost per model instead: at a 2.4x multiplier, "No way I would use this over opus 5.5".

The remaining comparisons are single voices, not a pattern. One developer says what fixed their context problem was a CLAUDE.md and AGENTS.md in every repo, read first every session, and calls Kiro's built-in memory "a nice step on from that" rather than a replacement. Another says they never dared to use the IDE and works in the CLI.

Nothing here was framed as a switch away with a reason beyond these: no one describes leaving for another tool, and no one names a tool they moved to Kiro from.

Locating code is the biggest theme among the questions, with 3 people. One asks how "retrieval accuracy" holds up when an entire codebase and spec go into a single pass; another asks whether a whole-codebase pass still misses files after an edit; a third wants to know how the tool finds the right file from an error screenshot, since so many real bug reports are a screenshot and "it broke".

The fix judge draws two questions: which product — Kiro, Kiro CLI or Kiro Crew — uses the self-improving feature, and whether the team will share a rejected fix that looked plausible along with the check that caught it.

When things arrive is the other question at that level: one asks when Kiro is coming to JetBrains now that it implements ACP, and another when the CLI gets OTEL support. Single people ask what the local sandbox in the IDE is actually for, whether data can be written to memory by hand, and where the jump over Sonnet 5 shows up, in speed or in code review.

The ask for Haiku 5.5 is consistent: it is described as "faster, better and cheaper", wanted as a lower-cost option alongside the current lineup, and at minimum swapped in for Haiku 4.5. One person making the request pairs it with a question about where the GPT 6 lineup is, and another names DeepSeek 4.1 flash as the kind of cheap, fast model they want.

The testing wish, from two people replying to the account, is about how fixes get accepted: keep a negative fixture where a boolean's value is the string "false", because a generic truthiness conversion turns it into true in some languages, and run each candidate fix across several seeds so a gain that shows on only one run is exposed by the median.

The rest are single voices. One wants frontier alternatives to Anthropic and OpenAI models or says they are gone; one wants repository-scoped memory to show which remembered correction shaped a suggestion so stale advice can be pruned; and one asks, playfully, for a haunted-house demo where the Wi-Fi works but DNS doesn't.