AWS Tackles Agent Drift With Bedrock AgentCore Optimization
AWS has introduced AgentCore Optimization in preview to automate prompt updates and A/B testing, alongside a new desktop AI assistant called Amazon Quick.
On May 4, 2026, AWS introduced the preview of AgentCore Optimization, a suite of performance management tools built into the Amazon Bedrock AgentCore platform. Designed to combat agent drift as models and user behaviors evolve, the toolset automates prompt tuning, batch evaluation, and A/B testing for production workloads. This infrastructure release follows the April 28 launch of Amazon Quick, a proactive macOS and Windows desktop AI assistant targeting enterprise workflows.
Managing Agent Drift
AgentCore Optimization establishes a performance loop by analyzing production traces from CloudWatch Logs. It identifies specific failure patterns, such as declining goal success rates or tool selection accuracy, and generates recommended adjustments to system prompts and tool descriptions.
Developers must approve all suggested changes. Before deploying updates, the platform uses batch evaluation to validate recommended modifications against pre-defined test datasets. If you spend time evaluating and testing AI agents, this feature automates the regression testing phase of prompt engineering.
To safely roll out changes, the AgentCore Gateway splits production traffic between the existing control agent and the updated treatment version. This provides statistical significance reporting on performance metrics. The platform also introduces Configuration Bundles, which act as versioned, immutable snapshots of agent configurations. These bundles capture prompts, model IDs, and tool descriptions together, allowing developers to update agent behavior without redeploying underlying application code.
Desktop Integration With Amazon Quick
While AgentCore handles backend orchestration, Amazon Quick operates as a proactive AI assistant running natively on macOS and Windows. The application builds a personal knowledge graph in the background based on a user’s local files and work context.
Amazon Quick connects directly to local applications and cloud services, including Google Workspace, Microsoft 365, Slack, Zoom, Salesforce, Airtable, and Dropbox. Instead of waiting for manual queries, the assistant surfaces proactive insights through OS-level notifications. It identifies calendar conflicts, drafts email replies, and summarizes meeting notes based on cross-application context.
A preview feature allows non-technical users to build custom internal tools. By describing a desired application in natural language, users can prompt Amazon Quick to connect live data sources and generate interactive applications, such as a CRM pipeline review dashboard.
Infrastructure and Pricing
AgentCore Optimization requires integration with AgentCore Observability and Evaluations. The feature is currently available in AWS Regions that support the core evaluations toolset.
Amazon Quick introduces a distinct pricing model designed to reduce friction for end users. The service allows registration with personal or corporate email addresses, removing the requirement for a dedicated AWS account.
| Plan Tier | Target Audience | Key Capabilities |
|---|---|---|
| Free and Plus | Individuals | Local file access, app integration, document generation |
| Professional and Enterprise | Organizations | Governance controls, agentic business intelligence |
The rollout of these tools aligns with broader infrastructure upgrades, including AWS’s recent partnership with Meta to support agentic workloads on Graviton processors. Early adopters of the new tools include GoDaddy, AstraZeneca, BMW, and Southwest Airlines.
If you manage production Bedrock deployments, integrating Configuration Bundles into your deployment process will help isolate behavioral prompt changes from your core application logic.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Run Multi-Tenant Agents With Dynamic Workflows
Learn how to use Cloudflare Dynamic Workflows to execute durable, stateful operations for multi-tenant applications and long-running AI agents.
ChatGPT Images 2.0 Adds Multilingual Text and Thinking Mode
OpenAI released ChatGPT Images 2.0 with the gpt-image-2 model, adding agentic web search, 2K resolution, and non-Latin script rendering capabilities.
ComfyUI Reaches $500M Valuation to Scale Node-Based GenAI
Comfy Org has secured $30 million in Series B funding led by Craft Ventures to scale its node-based AI generation platform as the control layer for creators.
Google Photos Adds 3D-Aware Re-composition via Auto Frame
Google has updated the Auto frame feature in Google Photos with a 3D-aware generative diffusion model to allow post-capture camera angle adjustments.
AWS SageMaker adds NVIDIA Blackwell G7e inference instances
Amazon SageMaker AI now offers G7e instances on NVIDIA RTX PRO 6000 Blackwell GPUs, with 96GB memory and 2.3x faster inference over G6e.