AWS adds persistent AI bots to manage recurring development work

Developers can now give an AI bot a continuing responsibility—such as checking failed builds—instead of restarting the assignment with each conversation. AWS launched Kiro bots in its Kiro Crew development workspace on November 5, making the feature generally available. AWS says the bots can remember instructions, supervise other agents and surface decisions that need human attention.

The launch adds a persistent coordinator to an existing workspace. AWS describes Kiro Crew as open source, with a desktop app, web dashboard and terminal interface. Crew already supports work across sessions, schedules and subagents—additional AI workers that take on parts of a larger job.

An assignment that survives the next session

Each bot has its own name, ongoing conversation, notes and workspace. Its memory belongs to its assignment, so deployment decisions can stay separate from a research bot’s context. Users can inspect, change or delete what it remembers. AWS describes saved context and corrections, not retraining of the underlying AI model.

According to AWS, a bot can launch multiple worker sessions, watch their progress, correct deviations and combine their results. If one branch needs a human decision, it can ask while continuing work that does not depend on the answer. Developers can return to the bot’s conversation to redirect the assignment without reopening every worker session.

Schedules can be described conversationally, including overnight failure checks or weekly change summaries. AWS offers several possible assignments; these are configurable workflows, not ready-made replacements for entire engineering functions:

• Triage incoming issues, group them and identify duplicates.
• Review proposed code changes against earlier engineering decisions.
• Check tests and documentation before a release.

A job title does not grant unrestricted access

Developers still define the job, connect the relevant systems and decide what counts as completion. Permissions can differ for each bot: a researcher might only read files, while a deployment bot could require approval before changing anything. AWS says adding tools requires user approval, and users can adjust each bot’s boundaries later.

The coordinator and its workers can also use different models. A team could assign planning to a more capable model and repetitive checks to a cheaper one. Kiro’s catalog includes OpenAI and Anthropic models alongside open-weight alternatives, with availability and credit charges varying by model. AWS has not supplied measured savings for that arrangement.

Crew can run on a laptop, Mac Mini or Amazon EC2 cloud instance, according to AWS. That choice determines where the workspace operates; it does not mean every underlying AI model runs locally on the user’s hardware.

The subscription is not the cost of a finished job

VentureBeat lists individual Pro access at $20 monthly with 1,000 credits. Teams of at least two pay $20 per seat monthly, with each seat contributing 500 credits to a shared pool. Additional credits cost $0.04 each.

A credit is not a completed task. Consumption depends on task complexity and model selection, so the monthly subscription alone cannot establish the cost of a continuously scheduled bot. More frequent checks, larger repositories and repeated attempts can all change the economics of unattended work.

The launch also leaves the central productivity claim unmeasured. VentureBeat reports that AWS supplied no comparative benchmarks, quantified productivity results or measured mistake rate. The capabilities are available, but how much human review and correction they require—and whether that reduces coordination work—remains unresolved.