A preview feature enters transition
AWS has stopped customers creating new batch or recurring jobs in Amazon Bedrock AgentCore Insights from 16 October 2026. Existing recurring jobs will continue to process sessions and produce reports until 30 October, after which they will also stop. Historical batch results and recurring reports remain available, and AWS says a redesigned Insights experience is coming soon.
The notice affects a public-preview capability, so change is expected, but the dates create an immediate operational task for teams that built evaluation or incident routines around it. Insights analyses agent sessions to identify failure patterns, extract user intents and summarise execution behaviour. Losing new analysis jobs, even temporarily, can leave a gap between raw traces and the clustered explanations operators use to prioritise fixes.
What Insights has been doing
The service goes beyond a single pass score. Its failure analysis can group tool errors, hallucinations, reasoning defects and repetitive behaviour, trace likely root causes to spans and propose remediation. User-intent extraction clusters the jobs people were trying to complete, while execution summaries reveal common approaches and outcomes. Across many sessions, those results can show whether one bug is isolated or represents a recurring operational pattern.
Previously, teams could request a one-time analysis over a time range or attach a daily, weekly or monthly clustering schedule to online evaluation. The pause means a new investigation cannot simply create a batch job, and a new production service cannot establish a recurring report. Existing schedules have a short runway, but relying on them until the final day would leave little time to test an alternative.
An interim observability plan
Teams should first inventory every workflow that consumes Insights output, including dashboards, tickets and release gates. Raw agent traces, tool results and evaluation scores should remain available independently of the generated report. Operators can then reproduce essential triage with deterministic queries: tool-error rates, approval failures, latency spikes, repeated loops and unsuccessful task outcomes can often be detected without a model-based clusterer.
Where semantic grouping is necessary, a temporary internal pipeline can work on a controlled export of sanitised traces. It should record the model and prompt used for classification, maintain links back to source evidence and avoid presenting automated root-cause suggestions as confirmed findings. The aim is continuity, not hurriedly rebuilding every preview feature.
Regional handling deserves attention
AWS also warns that Insights in Seoul, Mumbai and Singapore uses Bedrock global cross-region inference profiles. Requests may be routed to destination regions even where an organisation has not opted in or has service-control policies that deny inference there. Prompts and results may be stored in those destination regions for abuse detection. Australian and regional customers should treat that as a data-governance consideration, not a minor implementation detail.
Before preserving or exporting session data, teams need to classify what the traces contain. Agent logs may expose customer inputs, retrieved documents, tool arguments and operational identifiers. Retention, encryption and access controls should apply to the raw evidence as well as the derived report. A redesign is an opportunity to reassess whether the original configuration matched residency and privacy expectations.
Questions for the replacement
AWS has not provided a launch date or full design for the new experience in the cited notice. Customers should therefore avoid planning around unannounced capabilities. Useful clarification would include migration of schedules, API compatibility, supported regions, historical-report retention and whether the redesigned system changes how session content is stored or routed.
The replacement should also make uncertainty visible. Failure clustering can accelerate investigation, but similar traces do not always share one cause. Operators need access to the underlying spans, the ability to correct groupings and stable identifiers that allow a finding to be followed through remediation and regression testing.
A reminder about preview dependencies
Preview services are valuable for learning, yet they should not become the only path to evidence needed for safe operation. Production agent platforms need portable logging, clear evaluation criteria and an incident process that continues when an optional analysis layer changes. That foundation also makes it easier to compare the redesigned Insights feature when it arrives.
The short-term action is straightforward: stop assuming new jobs can be created, export the configuration and reports teams are entitled to retain, verify what existing schedules will do through 30 October and establish a bounded interim workflow. That work preserves continuity without pretending the old preview or its successor offers a complete assurance system.
Avoiding a blind period
Release owners should nominate a person responsible for the transition rather than assuming the preview will resume before existing schedules stop. A simple register can capture each job, its consumers, report frequency, data region and replacement control. During the gap, incident reviews should explicitly note which Insights-derived evidence is unavailable. That prevents missing analysis from being mistaken for an absence of failures. When AWS publishes the redesigned feature, teams can compare its outputs against the preserved raw traces and interim classifications before relying on it for a production decision.