In-depth analysis · About 4 min read
Tamnoon describes how Tami separates cloud investigation from approved repairs
An October 14 AWS case study explains four specialized agents, conservative safety scoring and controlled execution; reported improvements are customer-case claims.
- Source published
- Published here
- Source checked
Tamnoon and AWS have detailed the architecture behind Tami, a cloud remediation system built on Amazon Bedrock. This is a follow-up to their October 14 case study, rather than a new launch today. The useful distinction is between finding an exposure and authorizing a production change: the design places several checks between those activities.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Four agents divide investigation, safety assessment, repair planning and code generation. Only findings rated SAFE proceed to the remediation agent.
- Tamnoon reports up to 90% lower critical exposure within 90 days and zero production issues in its experience; these are vendor-reported outcomes, not an independent guarantee.
From alerts to a conservative decision
Tami sits downstream of CNAPP and cloud-native detection tools, normalizing their findings rather than replacing the detection layer. An investigation agent prepares read-only checks and a plan. A separate safety agent combines automated results with analyst evidence and classifies a proposed repair as SAFE, RISKY or AWAITING DATA, with uncertainty causing a downgrade. The remediation agent handles SAFE findings and orders changes around backups, verification and rollback. A fourth agent generates code after input screening; each artifact then needs a separate reviewer’s approval, with regeneration when necessary. This separation makes the decision to change infrastructure explicit.
Execution has a different trust boundary
According to the case study, investigation is restricted to reading, while the write path carries its own backup and recovery steps. The AI agents do not hold cloud credentials: an isolated execution environment uses short-lived, tenant-scoped credentials under customer-controlled trust. Approval workflows and audit trails remain part of governance. Generated output also goes through review and regeneration, with human involvement when the system cannot establish an accurate result. These layers reduce reliance on a prompt behaving correctly; they do not prove that every possible unsafe action has been eliminated.
A specific AWS implementation, not a universal promise
Bedrock supplies model access through one integration, with different tasks able to use different models. The described workloads invoke Claude through Bedrock, with versions pinned for reviewed changes. The implementation uses IAM-scoped roles, private networking, queues and operational alarms. Its cross-Region inference stays within the US geographic boundary described by the authors; that is not a claim of single-Region processing. EKS hosts persistent services while Lambda and workflow services process events. AgentCore and Guardrails support controlled actions and input/output checks, but the architecture still depends on correctly configured permissions and operations.
Read the result figures with their scope attached
The article reports more than 14 million alerts analyzed across Tamnoon’s customer base. Its results table states up to a 90% reduction in critical exposure in 90 days, up to 100-fold remediation efficiency through root-cause grouping, and more than 13,000 engineering hours saved per week across customers. It also reports zero production issues across millions of workloads. Those figures come from Tamnoon’s research and experience, with different denominators; they are neither one controlled test nor a promised outcome for a new customer. A team evaluating this pattern should examine its own approval policy, failure handling and rollback evidence.
Why it matters
For enterprise teams, the transferable idea is to make permission boundaries, evidence and rollback part of the execution system. Choosing a capable model alone does not establish whether a proposed cloud change is acceptable.
Limits and availability
The article does not provide independent incident verification, a reproducible comparative benchmark or a public subscription price. Its deployment and data-boundary statements describe Tamnoon’s implementation, not every Bedrock application or a personal Claude subscription.
Sources and evidence
Source published: Oct 14, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.