Based on AWS's Kiro Crew documentation as published on November 15, 2026. I have not tested any of this. Documentation pages change; I'll update this note if AWS answers any of the open questions below.
AWS made Kiro bots generally available on November 5: named, long-running agents with their own memory, each set up for a specific part of your work. The documentation is better than I expected, and it is worth saying what it gets right before asking what it leaves open.
What the docs state
Crew's security page says its controls are "enforced at the runtime boundary rather than relying on prompt instructions alone." Each tool call passes through a numbered series of checks, in order, and denied commands are enforced at Crew's own tool gate, not inside the agent's configuration.
For bots specifically, a bot asks for your approval before changing its own permissions, and permission changes "apply to running workers." One bot cannot read another bot's memory or control its workers. That is the right shape: the control sits at the action, outside the agent.
What a standing agent still adds
A session agent is born, acts and ends. A bot keeps a name, memory and a schedule across weeks. The pages I read leave four things open.
- Approvals over time. Scheduled activity "can run when the bot's chat is closed, while the Kiro Crew gateway is running." Approval requests expire if no one acts on them. What an approval means for a run that nobody is watching, and whether one carries into a later session, is not stated for bots.
- Memory as an input. Memory is meant to shape later work: the launch post says that if a bot chases an expected failure, you can correct it and future runs will apply the distinction. You can view, edit or delete what a bot remembers, and its workers share the same memory. The docs also say: "Neither memory isolation nor the separate working tree is a filesystem permission boundary." How notes get written, and whether each note records where it came from, is not stated. If a note can come from content the bot read rather than from you, it would carry into later runs the same way your corrections do.
- Revocation reaching live work. Deleting a bot "unbinds it from new sessions," and its workspace, memory and past transcripts are kept. What happens to workers and schedules already running is not stated. The separate "fresh start" action is, in the docs' words, "not a way to stop all workers or schedules."
- Identity. The bot's profile holds "the bot's identity," but which credentials a bot acts under when it calls a tool is not stated on these pages.
A control governs only the calls that pass through it
One caveat on Crew's security page shows why the boundary matters. It applies to Crew generally, not only to bots: with the Claude Code harness, "a tool already pre-approved in Claude's own settings bypasses Crew's approval flow," and Crew's deny rules and audit log do not observe that call. AWS documents this openly, which is to its credit. It is also the general point: an enforcement layer covers only the paths that run through it.
For any persistent agent, I would ask: what does an approval cover when no one is present, what can memory change later, and does revoking the agent stop what it already started?
A related point, in narrower form: a valid credential alone does not establish that an action was authorized (Signing Is Not Authorization).
Views are my own.
Sources (each read November 15, 2026):
- Kiro, "Introducing bots in Kiro Crew" (Nov 5, 2026): https://kiro.dev/blog/crew-bots/
- Kiro docs, How bots work (updated Nov 11, 2026): https://kiro.dev/docs/crew/bots/how-bots-work/
- Kiro docs, Manage bots (updated Nov 6, 2026): https://kiro.dev/docs/crew/bots/manage-bots/
- Kiro docs, Crew security (updated Oct 12, 2026): https://kiro.dev/docs/crew/security/
- Saleme, "Signing Is Not Authorization: Claim-Level Negative Vectors for Agent-Payment Receipts," Zenodo, v1.0, 2026-07-24: https://doi.org/10.5281/zenodo.21535452
Related reading
Your agent asked for approval once. What did it actually approve?What was the signer allowed to sign?
Your Agent Has a Valid Token. Should This Call Still Be Blocked?