AWS hands Kiro bots a memory and a to-do list

Amazon Web Services rolled out Kiro bots on Thursday, planting persistent AI agents inside its Kiro Crew development workspace. Now generally available, the bots are built to babysit recurring jobs, boss around other agents and hold onto what developers decide along the way. AWS's pitch: stop re-explaining yourself every time work picks back up.

Each bot gets its own name, a running conversation, a set of notes and a dedicated workspace, so developers can jump back in to redirect it or scan findings without reopening every worker session one by one. According to AWS, a bot can spin up several sessions, track their progress, correct drift and stitch the output together. When something needs a human call, the bot asks for it while pushing ahead on tasks that don't depend on the answer. Schedules are set just by talking: investigate overnight build failures, or assemble a weekly summary of changes. An activity log catches routine work while the conversation surfaces what actually needs eyes.

AWS sketches four sample gigs, including sorting incoming issues, flagging duplicates, watching build pipelines for fresh failures, reviewing proposed changes against earlier engineering decisions, and verifying tests and documentation before a release. Those are configurable workflows, not four autonomous replacements for the humans who do them — developers still define the job, hand over system access and decide what finished looks like. One example bot watches GitHub, groups issues by topic and shows backlog shifts on a custom dashboard.

Authority stays fenced in. AWS says a research-style bot can read files but not edit them, a deployment bot has to ask before changing anything, and a review bot can comment and request revisions but can't approve or merge code. Adding tools takes user approval. Memory is tied to each bot's assignment — deployment decisions stay separate from whatever a research bot absorbs — and users can inspect, revise or delete what's remembered. Notably, AWS doesn't claim the bots retrain any underlying models; the sales pitch is simply fewer repeated explanations.

Model choice is another dial. A coordinating bot and its workers can run different models, so teams can save a heavyweight for planning and hand repetitive checks to something cheaper. Kiro's catalog includes OpenAI and Anthropic models plus open-weight options, with availability and credit charges varying by model. Bots run through Crew on a laptop, a Mac Mini or an Amazon EC2 instance — though local hardware doesn't mean local models.

AWS is walking into a crowded ring. Cursor Automations starts at $30/month for Premium and keeps memory between runs that users can edit or delete, while Anthropic's Claude Code scheduled tasks — still tagged a research preview — lean on repository and connector configuration. OpenAI's scheduled work splits between local desktop tasks and web workflows. Kiro's plans start free with 50 credits, then Pro at $20/month with 1,000 credits; higher tiers run $40, $100 and $200, with extra credits at $0.04 each. Prices were checked November 5, 2026. AWS's real bet is the bundle: one continuing identity per responsibility, with its own memory, limits and a supervisor keeping score.