AWS's Kiro adds bots to Kiro Crew: long-running agents with memory that supervise parallel coding sessions

Kiro, the agentic development tool from Amazon Web Services, announced on November 5 that "bots" are now generally available in Kiro Crew, its environment for running several coding agents at once. A bot is a named agent that stays assigned to one area of work over time, rather than an assistant opened for a single task and then closed.

Agents that keep a memory and a goal

According to Kiro's announcement, each bot lives in a single continuing chat with the developer. It keeps its own notes on projects, decisions and open items, which the user can inspect, edit or delete, and that memory is isolated from other bots. Background and scheduled activity is written to a work log, and bots can maintain dashboards summarising the state of their work.

A bot is given a goal with milestones and works toward it on its own initiative. Developers can set schedules in plain language, such as checking for overnight build failures or drafting weekly release notes; the blog also shows a heartbeat option that wakes a bot at a fixed interval, with a cap on daily runs. Kiro's documentation positions bots as the choice for ongoing ownership of a goal, distinct from reusable custom agents, one-off scheduled jobs or in-session subagents.

Coordinating worker agents

For longer jobs, a bot launches parallel worker sessions, monitors them, steers those that drift and reports results in its chat. It can pause to ask the developer for a decision while unaffected work continues. Each bot only gets the tools and permissions its role needs; actions such as deployments wait for sign-off, and requests for extra tools require approval. Users can pick one model for the bot and another for its workers, and run bots on a laptop, a Mac Mini or an EC2 instance. Kiro's examples include bots for issue intake, build health, code review and release readiness.

Costs and oversight still open

Neither Kiro's post nor the trade coverage gives pricing for bots; the blog only shows credits consumed per result. Analysts quoted by InfoWorld were cautious. HyperFrame Research's Stephanie Walter advised starting with one bounded process and measuring correctness, review time and operating cost, noting that conventional automation may be cheaper for some recurring jobs. HFS Research's Ashish Chaturvedi warned that auditing decisions spread over weeks of sessions could get harder and that routine approvals may lead to permission creep. IDC's Dave McCarthy described a "coordination tax", as parallel retries and tool calls can quickly inflate token and compute spending.