AWS Releases AgentCore Harness for Production AI Agents with Two-API Setup

AWS Releases AgentCore Harness for Production AI Agents

Amazon Web Services made its AgentCore harness generally available today, eight months after its April preview launch. The service reduces production AI agent deployment to two API calls: CreateHarness to define an agent and InvokeHarness to run it.

The harness addresses what AWS identifies as the core friction in agent development: not the agent loop itself, but the surrounding infrastructure. Teams typically spend significant time on provisioning sandboxed compute, configuring storage and secrets, implementing observability, and managing concurrency when moving from laptop prototypes to production systems.

Technical Architecture

AgentCore harness provides six managed primitives: Runtime (isolated execution environment with filesystem and shell), Memory (conversation and user state), Gateway (tool connection layer), Browser (web interaction sandbox), Identity (credential management), and Observability (automatic CloudWatch tracing).

Each agent runs in an isolated microVM environment. Built-in tools include shell command execution and file operations without explicit configuration.

Multi-Provider Model Support

The service supports model switching mid-session while preserving conversation context. Available providers include:

• bedrock: Any model on Amazon Bedrock (Claude, Nova, Llama, DeepSeek, Qwen, Kimi, MiniMax, Cohere, Mistral, GPT-4o and GPT-4o-mini)
• openAi: Direct OpenAI API access
• gemini: Google Gemini models
• liteLlm: Third-party providers via LiteLLM (Anthropic direct, Cohere, Mistral, Vertex, Azure OpenAI)

According to AWS, teams can use Claude Opus for planning, switch to GPT-4o for code generation, then switch to Gemini for summarization within a single session. API keys are stored in AgentCore Identity's token vault; agents never access raw credentials.

Tool Integration

Five tool types are supported via configuration:

1. agentcore_gateway: Reference an AgentCore Gateway by ARN for OpenAPI, Smithy, Lambda, or MCP targets with IAM/JWT auth
2. remote_mcp: Direct connection to MCP servers by URL
3. agentcore_browser: Full browser sandbox for web interaction
4. agentcore_code_interpreter: Sandboxed Python and Node.js execution
5. inline_function: Tool schema for human-in-the-loop approvals or client-side execution

Tools can be overridden per invocation using the allowed_tools parameter.

Managed Memory

The GA release includes automatic memory provisioning. When the memory parameter is omitted on CreateHarness, the service provisions a managed Memory resource with SEMANTIC and SUMMARIZATION strategies, 30-day event expiry, AWS-owned encryption, and multi-tenant isolation keyed on actorId.

Teams can disable memory for stateless agents or attach existing AgentCore Memory resources by ARN.

What This Means

AgentCore harness standardizes production agent infrastructure on AWS, eliminating repetitive plumbing work for teams deploying multiple agents. The mid-session model switching capability addresses a specific pain point: testing price-performance tradeoffs or responding to model regressions without restarting conversations. For organizations already using AWS infrastructure, the managed approach removes container orchestration and reduces time-to-production. The catch: it's AWS-specific, and teams need to evaluate whether the abstraction layer matches their control requirements for production systems.