At the Bogotá Summit, AWS showcased Bedrock AgentCore, Fault Injection Service, Zonal Shift, and a set of post-quantum security guidelines.

Why AI agents are now a core building block

Amazon Quick Desktop, demonstrated at the event, runs locally, crawls a user’s professional environment, and builds a knowledge graph of relevant documents, contacts, and workflows. That front-end illustrates AWS’s push to make agents the default integration point for cloud workloads.

Bedrock AgentCore is the platform AWS is positioning to tame the growing agent ecosystem. It bundles three functions under the “Harness” concept:

  • Orchestration – define how multiple agents talk to each other and to downstream services.
  • Observability – collect metrics, logs, and traces so operators can see which agent handles each request.
  • Control limits – set quotas and safety guards that prevent runaway execution or data leakage.

Developers use the service to spin up specialized agents without building the plumbing from scratch. By exposing a standard API for these three functions, AWS wants agent development to become a repeatable pattern rather than a niche experiment.

Resilience moves beyond “keep the lights on”

High-availability promises have shifted from simple uptime percentages to measurable error budgets backed by real-time observability. At the summit, Yuno—a regional SaaS provider—showed how it sustains a 99.95 % SLA by combining canary deployments with chaos engineering.

Two new tools stole the spotlight:

  • Fault Injection Service – lets teams inject latency, errors, or node failures into a live environment to verify monitoring and remediation logic before a real outage.
  • Zonal Shift – a traffic-routing primitive that moves inbound requests from a failing Availability Zone to healthy zones without touching application code.

AWS walked through a component-by-component mitigation checklist:

  • Amazon Aurora – enable regional failover with Aurora Global Database.
  • Amazon MSK – use MSK Replicator to keep topic offsets in sync across clusters.
  • Load balancers and EKS – apply Zonal Shift on Application Load Balancers and on Amazon Elastic Kubernetes Service clusters.

For migrations that can’t tolerate downtime, the summit promoted a Blue/Green workflow: spin up a parallel cluster, drive 1 % of traffic through ArgoCD-managed releases, and cut over the full load only after validation. The approach leans on the same observability data that fuels error-budget calculations.

Getting ready for quantum-level attacks

AWS reminded attendees that quantum computers, while still experimental, will eventually break widely used public-key algorithms such as RSA and ECC. Amazon Braket now offers access to real quantum processors and high-fidelity simulators, giving security teams a sandbox to test cryptographic primitives.

The security track urged organizations to start a post-quantum cryptography (PQC) migration roadmap now:

  1. Inventory every cryptographic asset – TLS certificates, VPN keys, data-at-rest encryption modules.
  2. Apply hybrid controls – run a classical algorithm in parallel with a candidate PQC algorithm until the latter proves safe.
  3. Plan for hardware acceleration – watch emerging PQC chips that could be integrated via AWS Nitro or custom FPGA instances.

The guidance stops short of mandating any specific algorithm.

Declarative security perimeters in practice

Traditional cloud security has focused on identity (who can log in) and network (which subnets can talk). AWS used Bancolombia’s Control Tower deployment as a template for a four-layer “perimeter” model:

  • Identity perimeter – Service Control Policies (SCPs) that restrict what actions IAM roles can invoke across the organization.
  • Resource perimeter – Resource Control Policies (RCPs) that tag and guard specific services, such as limiting S3 bucket creation to approved regions.
  • Network perimeter – VPC Endpoints and the newer VPC Lattice keep traffic on the AWS backbone and avoid the public internet.
  • Declarative policy enforcement – CloudFormation Hooks and Lambda functions automatically remediate drift, for example by re-applying encryption settings when a bucket is created without server-side encryption.

The idea is to codify every security decision so drift is caught and corrected without manual ticket churn.

Real-world AI at scale: ADRES Colombia

Konkretną ilustracją paradygmatu agentowego jest ADRES, kolumbijski procesor roszczeń zdrowotnych. Ich stos technologiczny łączy Amazon Nova (zarządzaną usługę wnioskowania) z SageMaker, orkiestrowaną za pomocą architektury wieloagentowej:

  • Wyspecjalizowani agenci analizują faktury, wyodrębniają kody kliniczne i równolegle weryfikują uprawnienia.
  • Warstwy wyjaśnialnej sztucznej inteligencji (XAI) dołączają metadane pochodzenia, dzięki czemu audytorzy mogą prześledzić decyzję aż do oryginalnego dokumentu.
  • Kroki typu Human-in-the-loop (HITL) wstrzymują zautomatyzowany przepływ pracy, aby audytor medyczny mógł wydać ostateczną zgodę przed dokonaniem płatności.

Kontrargument: złożoność i koszty

Wszystkie te nowe prymitywy obiecują automatyzację, ale jednocześnie podnoszą poprzeczkę w zakresie wiedzy operacyjnej. Wdrażanie Fault Injection Service i Zonal Shift wymaga dojrzałego potoku CI/CD oraz niezawodnej telemetrii; w przeciwnym razie mechanizmy zabezpieczające staną się źródłem szumu. Utrzymywanie hybrydowych stosów PQC dodaje narzut kryptograficzny, który może wpływać na obciążenia wrażliwe na opóźnienia. Mniejsze firmy mogą uznać początkową inwestycję za zaporową, dopóki narzędzia nie dojrzeją, a ceny się nie ustabilizują.