On November 5, 2026, Amazon Web Services made Kiro bots generally available, persistent named agents built to coordinate parallel development sessions. AWS’s bet is to hand supervision to agents with their own memory, narrow permissions and recurring schedules, rather than to a general-purpose assistant that must sort through its own toolset.
🔑 Key takeaways
- Kiro bots are named, persistent agents with their own notes, work log and memory, available inside Kiro Crew.
- Each bot receives only the tools it needs; any additional tool request takes effect only after approval.
- Four typical assignments are offered: ticket triage, build health, merge review and release preparation.
- No separate per-bot price was announced; Pro costs $20 per month for 1,000 credits, and a credit does not equal a completed task.
- AWS publishes no comparative benchmark, error rate or quantified productivity measure.
A coordination problem created by parallel work
The launch addresses a problem that appears as soon as developers run several agents at once: they must spot blocked sessions, reconcile outputs and restate requirements after every interruption. AWS wants to move that coordination load onto bots, each responsible for a distinct task and able to supervise other agents.
The project has internal roots. Kiro Crew launched on August 4, 2026 as an agentic orchestrator active day and night. It descended from MeshClaw, an internal tool built by three Amazon engineers to start a task, step away, and return to a result ready for review while running several jobs at once. According to Kathy Kam, director of agentic engineering at AWS, that internal tool reached 39,000 Amazon developers and nearly 500 contributors within six months.
“Each bot is an agent tuned with the right set of skills, tools and memory, so it can do its job as well as possible, as opposed to a generalized agent that has to sort through a set of tools to determine which to use.”
Kathy Kam, Director of Agentic Engineering, AWS
The naming choice is also telling. AWS had considered “crewmates” and “jams” before settling on “bots”, a term a spokesperson judged more descriptive of its lasting benefits. Kam summed up the promise in two words: less mental load for the developer, and an agent that grows more relevant as it is corrected.
Memory, permissions and schedules: the anatomy of a bot
A persistent identity
Each bot is a named agent with an avatar, a single durable conversation, its own notes, a workspace and a customizable dashboard. Its memory is tied to its function and does not mix with that of other bots; it can be viewed, edited or deleted. AWS states that learning comes from saved context, conversations and corrections. The announcement does not claim to retrain the underlying models.
Segmented permissions
The principle of least privilege sits at the core of the offer. A research bot can search the web and read files without modifying them. A deployment bot runs scripts but asks for approval before each change. A review bot comments and requests changes without approving or merging. Boundaries remain editable, and any extension of access must be validated.
The coordinator and its worker agents may also run different models: a more capable model for planning, and a faster, cheaper one for repetitive checks. The catalogue includes OpenAI and Anthropic offerings as well as open-weight alternatives, with availability and credit costs varying by option.
Conversational scheduling
Schedules are defined in plain language: overnight investigation of compilation failures, a weekly summary of changes, or continuous integration checks every 30 minutes. Scheduled activity feeds a work log, and findings, handoffs and questions needing intervention return to the chat. When a task requires a human decision, the bot flags it and keeps working on independent tasks.
The table below summarizes the four assignments AWS presented. The developer always defines the task, the access and the completion condition.
| Bot | Mission | Permissions | Checkpoint |
|---|---|---|---|
| Ticket triage | Labels, duplicate detection, reproduction steps | Read and draft, no closing | Human validation of closures |
| Build health | Continuous integration monitoring every 30 minutes | Read and diagnose | Alerts only on new failures |
| Merge review | Checks requests against past decisions | Comments and change requests | No approval or merge |
| Release preparation | Checks tests and release notes | Prepares deliverables | Human sign-off before production |
Dashboards, public data and the limits of the announcement
The bot maintains a dashboard of its activity. The “Issue Observatory” demo shows, for the kirodotdev/KiroCrew repository, 1,468 open issues, including 316 opened over seven days, along with 367 bugs against 695 feature requests. MCP and OAuth servers account for 74 open issues, of which 30 are new, or 41% of new issues: an example of a trend a monitoring bot can surface automatically.
These figures illustrate tracking capacity, not agent performance. The announcement offers no comparative benchmark, no measured productivity result and no error rate. The four assignments are configurable workflows, not functions already replaced autonomously. An enterprise evaluation should therefore measure review and correction time, work actually completed, unexpected actions and model consumption.
Pricing and competitive landscape
Kiro subscriptions cover Crew, the development environment, the command-line interface and the web product. The free tier includes 50 credits, Pro costs $20 per month for 1,000 credits, and higher individual tiers cost $40, $100 or $200. Extra credits cost $0.04. For teams of at least two people, AWS lists $20 per seat per month, with 500 credits per seat. AWS provides neither measured savings from cheaper worker agents nor a guaranteed monthly cost for a scheduled bot.
Competitors already offer scheduling, retained context and background execution. The table compares entry prices verified on November 5, 2026.
| Tool | Entry price | Scheduling | Models |
|---|---|---|---|
| Kiro | Pro, $20/month | Persistent bots, conversational schedules | Multi-provider announced |
| Cursor Automations | Premium, $30/month | Schedules and events, notes retained | OpenAI, Anthropic, Google |
| Claude Code | Claude Pro, $20/month | Scheduled cloud tasks, in research preview | Claude models only |
| ChatGPT/Codex | Plus, $20/month | Scheduled feature | OpenAI models only |
Kiro’s differentiation rests less on persistent memory, which rivals also offer, than on the combination of a durable conversation tied to one responsibility, its own authority, and a coordinator that supervises worker agents. This is a design difference, not proof of superiority.
“The difference is not the tool. It’s how teams work with it.”
Swami Sivasubramanian, Vice President of Agentic AI, AWS
That remark captures AWS’s thesis. On its agentic AI page, the group says internal teams that redesigned their processes around agents delivered, in 76 days with six engineers, a project initially planned for 30 developers, with median gains of 4.5 times. These results remain company claims, with no published methodology.
Conclusion: a bet on discipline more than on the tool
Kiro Crew and its bots convey a clear idea: the useful agent is the one whose scope, memory and rights are narrow and explicit. For a development environment vendor, this is also a way to contest scheduled automation, a ground where Cursor and Claude Code have already established positions.
Three scenarios stand out. In the first, independent measurements confirm a net reduction in review time, and bots become a standard. In the second, credit-based billing, whose consumption depends on task complexity and model choice, slows adoption of recurring tasks. In the third, teams keep responsibilities narrow, as AWS recommends, and expand only after reviewing results. Follow-up publications from AWS and feedback from engineering teams will be decisive.
Sources
- VentureBeat, AWS launches agentic AI Kiro bots
- AWS, agentic AI page
- Kiro, Crew bots launch post
- CodeSprint Consulting, Kiro bots
- The Letter Two, AWS Kiro bots and Kiro Crew
- SiliconANGLE, Kiro Crew launch, August 4, 2026
Cet article est publié à titre informatif et éducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute décision.
