Amazon Bedrock Adds Code Execution: Week of 16 June 2025

Amazon Bedrock just crossed a significant threshold this week, introducing inline code nodes that allow direct code execution within AI workflows. This isn't just another feature addition: it's a fundamental shift towards more programmable AI systems that could reshape how developers build complex applications.

The Big Moves

Amazon Bedrock Gets Programmable with Inline Code Nodes

Effective 19 June 2025, Amazon Bedrock introduced inline code nodes in preview, marking the platform's most significant capability enhancement this year. This feature allows developers to execute custom code directly within their AI workflows, moving far beyond simple data transformations into full programmatic control.

The implications are substantial. Previously, Bedrock flows were largely constrained to predefined operations and API calls. Now, developers can inject custom logic, perform complex data manipulations, integrate with external systems, and create sophisticated conditional workflows all within the same execution environment. This positions Bedrock as a serious contender for enterprise AI orchestration, competing directly with platforms like LangChain and custom-built workflow engines.

For existing Bedrock users, this opens up entirely new use cases. Consider a customer service application that can now dynamically query internal databases, perform custom calculations, and even trigger external APIs based on conversation context, all without leaving the Bedrock environment. The preview status means AWS is likely gathering feedback on performance, security boundaries, and execution limits before general availability.

Developers should start experimenting with this capability immediately. The preview period is typically when AWS refines features based on real-world usage, and early adopters often influence the final implementation. Expect to see execution time limits, memory constraints, and supported programming languages expanded based on community feedback.

Google Expands Vertex AI Model Garden with DeepSeek

Google quietly expanded Vertex AI's Model Garden on 16 June 2025, adding the DeepSeek API in preview mode. While this might seem like routine model addition, it signals Google's continued strategy of becoming the Switzerland of AI platforms, offering multiple model providers rather than pushing only their own Gemini models.

DeepSeek's inclusion is particularly interesting given their focus on reasoning-heavy tasks and mathematical problem-solving. This complements Google's existing model roster and provides developers with more specialised options for specific use cases. The preview status suggests Google is testing integration patterns and usage metrics before full rollout.

For Vertex AI users, this represents another option in an increasingly crowded model marketplace. The key question is pricing and performance compared to existing alternatives. Google's model garden strategy appears designed to prevent vendor lock-in concerns that might drive enterprise customers towards multi-cloud AI strategies.

Worth Watching

Bedrock Extends Flow Execution Durations

Also on 19 June, Amazon Bedrock introduced extended flow execution durations in preview. This seemingly minor capability enhancement actually addresses a significant limitation for complex AI workflows. Previously, Bedrock flows were constrained by relatively short execution windows, limiting their use for batch processing, complex analysis, or workflows requiring multiple external API calls.

The extended durations open up new possibilities for enterprise use cases like document processing pipelines, comprehensive data analysis workflows, and multi-step approval processes. While AWS hasn't specified exact duration limits, this change suggests they're positioning Bedrock for more substantial enterprise workloads rather than just quick conversational AI applications.

Quick Hits

• Mistral AI released Mistral Small 3.2 (20 June): Performance improvements to their compact model, though specific benchmarks weren't disclosed
• Replicate platform improvements (20 June): Enhanced navigation, updated API playground with prediction links, and expanded video format support for better user experience
• Replicate adds model metadata environment variables (18 June): New debugging capabilities with username, model name, Docker image URI, version ID, and deployment name accessible within containers
• Replicate optimises API responses (16 June): Reduced model metadata by 5KB per object, improving LLM performance and reducing response times, particularly for MCP server operations

The Week Ahead

Watch for AWS to announce general availability timelines for both Bedrock inline code nodes and extended execution durations. The preview feedback period typically runs 4-8 weeks, suggesting potential GA announcements in mid-July.

Google's expansion of Vertex AI model garden suggests more provider additions are coming. Keep an eye on their I/O announcements for hints about which models might join the platform next.

Replicate's recent optimisation push indicates they're preparing for increased usage, possibly ahead of a significant platform announcement or pricing changes.

The broader trend this week shows major providers focusing on developer experience improvements rather than headline-grabbing model releases. This suggests the market is maturing beyond pure capability competition towards platform usability and integration depth. For enterprise teams evaluating AI platforms, these operational improvements often matter more than benchmark scores.

Expect continued focus on workflow orchestration capabilities as providers recognise that most enterprise AI applications require complex, multi-step processes rather than simple model inference calls.