AWS Well-Architected Agent enters preview with AI-driven infrastructure recommendations

AWS released AWS Well-Architected Agent in public preview on October 1, 2026, bringing AI-powered infrastructure analysis and remediation guidance to teams managing cloud environments. The service reviews AWS workloads against the Well-Architected Framework’s best practices and delivers prioritized recommendations tailored to the business goals you declare during setup.

The agent addresses a practical friction point in cloud operations: knowing what to fix first when infrastructure sprawls across dozens of services and hundreds of resources. Traditional architecture reviews produce static reports that require manual prioritization and research into implementation steps. Well-Architected Agent automates both the analysis and the remediation path, correlating utilization metrics, resource configuration, and application topology to surface improvements across cost, security, performance, and resilience.

Setup requires creating an agent profile in the Well-Architected console. You choose which accounts and regions to monitor, which Framework pillars to prioritize, and what goals matter most within each pillar. The service then provisions a customer-managed IAM role with read access to resource configuration, utilization metrics, and application topology. Recommendations appear within 24 hours and refresh periodically as your environment changes.

The agent delivers three tiers of findings. Individual resource recommendations flag single-resource issues like undersized instances or missing encryption, often with cost impact estimates and staged remediation steps. Multi-resource findings span an application’s architecture and identify systemic patterns—say, a missing disaster recovery plan across a microservices deployment. Infrastructure as Code recommendations review your Terraform, CloudFormation, or CDK projects and suggest changes to align with Well-Architected patterns before deployment.

Remediation comes in three formats: console tutorials that walk through manual changes, updated IaC files you can review and apply, or AWS CLI commands for scripted fixes. Each option includes step-by-step instructions. The service also supports pre-deployment reviews: upload a .zip archive of your IaC project, choose a Well-Architected lens, and receive feedback before provisioning resources.

AWS states the agent checks workloads against best practices across more than 65 AWS services. The complete service list is not published, but the breadth suggests coverage extends beyond the core compute, storage, and networking services into databases, analytics, machine learning, and security tools. The agent can onboard workloads from any AWS commercial region, though recommendations themselves are generated in US East (N. Virginia), US East (Ohio), or US West (Oregon).

Access requires an AWS Support plan. AWS positions the agent as a successor to manual Well-Architected Tool reviews and frames it as intelligence built on the same Framework that has guided architecture decisions for years. The existing Well-Architected Tool remains available for teams that prefer manual evaluation or need custom lenses.

The service exposes a programmatic API, and AWS points to integrations with the AWS MCP server and plugins for AI coding tools. This opens a path for developers to request recommendations directly from their editor or CLI without switching to the console. It also makes it possible to automate remediation workflows—for example, opening pull requests with the agent’s IaC diffs or triggering approval workflows for high-impact changes.

AWS includes a disclaimer that recommendations are generated by AI and may contain errors or incomplete information. Customers are responsible for evaluating suggestions in their own context, and the service does not guarantee that every recommendation will improve every metric. This is a standard caution for generative systems, and it reflects the reality that architecture trade-offs depend on workload-specific constraints the agent may not fully observe.

The preview raises questions that AWS has not yet addressed. Pricing details are not published beyond the AWS Support plan requirement. The agent’s underlying model is not named. Recommendation refresh frequency is described only as “periodically,” leaving teams uncertain whether they will see daily updates or weekly batches. There is no published accuracy baseline or false positive rate, so early adopters will need to calibrate trust through their own testing.

Well-Architected Agent represents a practical application of AI to a historically manual, time-intensive process. Architecture reviews have always been valuable but expensive in engineering hours. Automating the correlation of metrics, config, and topology with a body of best practices compresses that timeline and lowers the barrier to acting on findings. Whether the recommendations prove accurate and actionable at scale will depend on how well the agent’s model handles the edge cases and context-specific trade-offs that define real-world infrastructure. The preview period will test that.

The “AWS Support plan” line is doing more work than it looks. The AWS WA Agent user guide’s access table lists Developer and Business tiers as No access, with entitlement starting at Business+ (2 profiles, 7 applications per profile) and rising to 10 profiles and 30 applications at Enterprise On-Ramp, Enterprise Support and Unified Operations, while accounts per profile stays at 100 across the entitled tiers. So “requires an AWS Support plan” understates the gate for anyone sitting on Business tier, and the profile and application caps are the real scoping constraint for a multi-account review. If the plan is to onboard several business units under one profile, the numbers to design around are those per-profile limits rather than the plan name.

The dollar-impact figures are the part I would want pinned down before a cost recommendation goes into a budget case. AWS’s related-services page says cost attribution uses Cost Explorer data, and that with Cost Explorer not enabled the agent estimates costs from public AWS pricing instead – list price, not your negotiated rates or measured spend. That distinction matters for the same reason a rightsizing finding needs a defensible cost basis: the recommendation can be directionally correct and its number still not reconcile with what you are actually billed. It would be worth knowing whether an EDP or private pricing agreement is reflected anywhere in that estimate before treating the figure as a savings target.

What the announcement leaves implicit is how much of the Framework the agent actually covers. The Well-Architected Framework’s own documentation still describes six pillars – operational excellence, security, reliability, performance efficiency, cost optimization and sustainability – while the AWS WA Agent user guide says the agent analyzes across four optimization pillars: cost optimization, security, resilience, and performance. Operational excellence and sustainability do not appear in that list, so a profile configured to prioritize “pillars” is really a four-of-six review rather than a full Framework assessment. That reads less as a defect than as a scoping note for anyone expecting the automated review to stand in for a complete Well-Architected review, and it would be useful to know whether the remaining two are roadmap or deliberately out of scope.

The access-model page is worth reading alongside the setup description, because the roles are yours rather than something the service provisions for you. The agent assumes an execution role in the profile account and then chains into a separate access role in each target account, with the profile ARN supplied as the external ID for confused-deputy prevention and the WellArchitectedAgentResourceScanning managed policy attached to each access role. Onboarding is described as opt-in per account, which matters if some teams want to stay outside the scope. My open question is the failure mode: when an access role in a target account is revoked or its trust policy drifts, does the next generation run surface a partial scan somewhere visible, or does that account simply go quiet in the results?

The refresh interval the post files as an open question is stated in AWS’s own documentation: the user guide says scheduled recommendations are refreshed on a weekly cadence, and the quotas page puts a seven-day cooldown between scheduled generation runs for the same profile. The limit next to it is the one that shapes a rollout – a maximum of 30 recommendations per scheduled or manual generation run, with no cap on total stored recommendations and later runs adding more – so a large estate gets findings throttled into weekly batches of 30 rather than one complete first pass. The middle tier carries its own caveat, with AWS describing application-level recommendations as a new format currently in beta and inviting feedback on their quality. For a team feeding these into a change process, does a weekly batch of 30 land faster or slower than your remediation capacity?