Item: Kiro
Author: GILD

GILD verdict

Choose Kiro if your team wants agent work planned as specs and bought and governed through AWS.

Require managed sign-ins from day one, and keep specs short enough that reviewers read them.

Kiro

AI Coding Agent by Amazon Web Services

+10 Mixed

Fit and limits

Who it fits

You want each feature or bug fix written up as requirements or a bug analysis, a design, and tasks before the agent writes code, kept in the repository for review.
You want conventions in steering files and hooks that apply to every member, with team steering pushed to machines through MDM.
You buy through AWS and want seats in the AWS console, IAM Identity Center or a direct Okta or Entra ID connection, and AWS billing.
Your engineers use Amazon Q Developer. Kiro CLI replaced Q Developer CLI, and AWS ends support for the Q Developer IDE plugins and paid subscriptions on April 30, 2027.
Your agent sessions run long. In AIMultiple's June 2026 test, Kiro CLI kept all 13 facts after compacting its context, while three other agents on Claude Sonnet 4.6 kept none.

Where it falls short

Specs add time and review load that one-line fixes and fast exploration do not repay. In one public evaluation, a small bug fix came back as 4 user stories with 16 acceptance criteria.
Credits scale with the model, and extra credits cost twice the plan rate, so spend needs a default model and an overage cap.
Kiro Web's cloud sessions run only in US East (N. Virginia), and the admin model list, MCP settings, and customer managed keys do not apply there.
Results vary run to run. Kiro CLI ranked 7th on AIMultiple's build tasks (benchmark updated Jun 29, 2026). On the one task AIMultiple broke down in detail, three reruns ended three different ways and averaged 24.4% on backend checks.
Engineers on personal GitHub, Google, or Builder ID sign-ins fall under individual terms, where AWS can use their content for service improvement unless they opt out.

Start on managed sign-ins
Create a Kiro profile in an AWS account and Region, connect IAM Identity Center, Okta, or Entra ID, and subscribe users or groups to a tier. Enterprise users' content is not used for service improvement and their telemetry is off. The Free tier is not available to them.
Set the shared defaults
Commit steering files and hooks, approve a model list and a default model, limit MCP to a registry, and decide whether web tools and API keys for headless CLI runs are allowed. Kiro IDE asks before writes, commands, and MCP calls until rules allow them. On Kiro Web, background workflows pause when a step needs input and stop at an iteration cap until someone grants 1 to 1,000 more.
Write a spec rule
Decide which changes get a feature spec, a Bugfix Spec, or plain chat, and how long a spec can run. Spec tasks cost more credits than chat prompts, and reviewers have to read what the agent writes before it builds.
Cap spend
Overages are off by default. Once enabled, they bill at US$0.04 per credit at month end, up to a per-account cap in Service Quotas. Assign tiers by observed usage. An engineer who uses more than 1,500 credits a month costs less on Pro+ than on Pro with extra credits.

GILD searched AWS case studies, practitioner posts and benchmarks and found no team-level rework figures for specs; tag pilot tickets spec or chat and compare review rounds per merged change.
Sonnet 5.5 costs 1.3 times Auto's credits, so run 20 tickets on each and pin Sonnet 5.5 only if accepted changes rise by more than 30%.
Before allowing Kiro Web, send it 10 tickets and count how many reach review without help, because the admin model list and MCP settings do not reach it yet.