Amazon Web Services introduced Kiro bots on Thursday, adding persistent agent memory and delegated workflow management to its agentic development toolkit. The new feature gives each bot its own durable memory and the ability to run parallel worker sessions independently, addressing what developers call "context rot" — the problem of agents forgetting instructions between sessions. AWS positions the update as a shift from delegating individual tasks to delegating entire processes, allowing software engineers to define workflows once rather than shepherd agents through each step.

Each Kiro bot functions as a named, long-running agent customized for a specific responsibility, such as monitoring a ticket queue or managing deployments. The bots retain their own memory through a notes file they read and write, which developers can edit directly to fix mistakes that persist across all tasks in flight. Bots work only with the tools and permissions assigned at creation — a research bot can search the web and read files but can't modify them, while a deployment bot can execute scripts and must request approval before each change. Each bot operates with separate memory, so lessons learned by the deployment bot about release processes don't mix with what the research bot discovers. Bots launch and monitor their own parallel sessions without requiring developers to track them individually or weave results together, and they can redirect sessions that drift from the plan while combining outputs when work finishes.

According to Kathy Kam, AWS Agentic AI director of engineering, agents become useful only when developers trust them to act independently, but when a session ends, nobody owns the work anymore. "Developers end up as the go-between, carrying context from one session to the next," Kam said. "Kiro bots take that job off their plate, so developers can now spend the day deciding what's worth building instead of babysitting sessions." Kam explained that without durable memory, context rot spreads quickly when parallel tasks each maintain their own drifting version of reality, making it feel like nagging an agent to repeat yesterday's instructions. She emphasized that trust comes from track record rather than agent intelligence: "A hierarchy of agents based purely on IQ is the wrong measure. What matters is what an agent is allowed to touch, and a Kiro bot earns trust by building a track record."

The report explains that as companies move toward software factories, agents can remember more information and context across a team than individual developers hold, but the human remains accountable while the agent executes the work. Software engineers set access levels for each bot, determining whether it requires approval before every change or runs with fewer checkpoints. Bots can request additional authority but cannot take it without approval, and while waiting for human decisions, they continue working on tasks that don't require input. By design, bots can't delegate work to other bots — AWS built this limitation to prevent context rot and keep bots independent. Each bot splits large jobs across parallel workstreams and consolidates results into one recommendation, while developers define boundaries and connections between bots so no decision moves between them without involvement. Over time, a developer's crew of bots reflects their actual development approach, with each bot carrying requirements and boundaries for its portion of work while the engineer decides how processes change and where judgment belongs. For organizations navigating the balance between automation speed and human oversight, this architecture suggests a future where control becomes more about designing guardrails than executing every step. The question isn't whether agents can outpace human judgment in narrow domains — it's whether teams can build trust frameworks that scale without creating new coordination costs that offset the efficiency gains.