Content tagged with "agentcore"
Stood up an agent in us-west-2 from two API calls with only a name and execution role, timed the idea-to-first-response, and proved context survives a mid-session swap from Claude to OpenAI.
Use the AgentCore CLI (Node.js) to deploy a Strands agent to Bedrock AgentCore with just agentcore create + deploy. Verified that session isolation enables conversation continuity across invocations.
Tested Stateful MCP's 3 capabilities with a DynamoDB-backed expense tracker. Verified data integrity on Elicitation decline/cancel and fallback behavior on Sampling failure to provide production design guidance.
Verified Agent Registry end-to-end: registry creation, record registration, approval workflow, and hybrid search via CLI. Measured semantic search accuracy, Japanese vs English query differences, and MCP endpoint consistency.
SAM deployment gotchas, local testing with LocalDurableTestRunner, and execution history tracking from hands-on experience. The qualified ARN requirement is the first hurdle.
Verified idempotency, DurableExecutionName, parallel callbacks, and timeout configuration hands-on. The common principle: design everything assuming replay will happen.
Deployed the official AWS fraud detection demo and verified 3 risk-score branches with suspend behavior. Confirmed zero compute charges during checkpoint-and-replay waits.
A deep dive into the AWS Security Blog's agentic AI security framework for financial services. Reorganizes the 7 design principles into three axes—permissions, traceability, and controls—with Bedrock AgentCore implementation guidance.
Hands-on verification of AgentCore Runtime's new managed session storage. Tested file restoration after stop/resume, pip/git workspace persistence, session isolation, and graceful shutdown behavior with real measurement data.
Combine Runtime, Memory, Gateway, and Evaluations into a single AgentCore CLI project to build a personalized tech trend advisor that remembers user expertise, searches the web for latest trends, and auto-measures response quality.
Hands-on verification of AgentCore CLI Evaluations with a custom LLM-as-a-Judge evaluator. Covers on-demand evaluation of agent traces, region constraints for the Evaluator CloudFormation resource, and trace indexing timing.
Hands-on verification of AgentCore CLI Gateway with an external MCP server target. Covers the critical project setup order, deployment-time connectivity checks, and the split between Gateway API and CloudFormation deployment.
Hands-on verification of AgentCore CLI Memory with SEMANTIC, USER_PREFERENCE, SUMMARIZATION, and EPISODIC strategies. Covers async extraction timing, session ID constraints, and the gap between local dev and deployed behavior.
Hands-on verification of the AgentCore CLI — from creating a Strands Agents project to local development, AWS deployment, and observability. CDK knowledge not required; initial deployment completes in about 90 seconds.
Deploy a bidirectional voice streaming agent using AgentCore Runtime's new WebRTC support with Nova Sonic. Key finding: TURN-only mode is strongly recommended for AgentCore Runtime deployments.
Hands-on verification of AgentCore Code Interpreter's S3 integration and PUBLIC mode. Covers direct pip install, S3 wheel offline installation, sys.path import, and cross-session persistence behavior.
Hands-on verification of AgentCore Runtime's new shell command API. Covers HTTP/2 streaming, in-session file persistence, cross-session microVM isolation, and concurrent execution behavior.
Deploy an AG-UI protocol server to Bedrock AgentCore Runtime and stream real-time agent responses via SSE events. Hands-on walkthrough with agentcore CLI Direct Code Deploy.
Hands-on verification of Bedrock AgentCore Policy: Cedar-based tool access control via Gateway, natural language policy generation, and default deny behavior validated with real API calls.
Hands-on verification of AgentCore Memory Record Streaming. Covers Kinesis push delivery, FULL_CONTENT vs METADATA_ONLY behavior, and async extraction events with real test data.
Verifying AgentCore Runtime's new stateful MCP features. Tested elicitation (server-initiated input), sampling (LLM generation requests), and progress notifications both locally and on AgentCore.
After overriding InvokeHarness's model from Claude to OpenAI gpt-oss, the gpt-oss model inherited the history and claimed to be Claude. Don't trust a model's self-report to tell which one is running.
AWS Agent Registry CLI uses skillMd (dict with inlineContent) while REST API uses skillMarkdown (string). descriptorType is AGENT_SKILLS, not SKILL.
AWS Agent Registry MCP records require reverse-DNS name format, 100-char description limit, and explicit schemaVersion. The error message doesn't tell you which field is wrong.
Setting us-east-1 in aws-targets.json doesn't help if AWS_REGION is set to ap-northeast-1. CDK under the hood picks up the env var. Not documented in CLI docs.
agentcore deploy fails with Unrecognized resource types: [AWS::BedrockAgentCore::Evaluator] in ap-northeast-1. Switch to us-east-1.
agentcore deploy fails with CloudFormation ROLLBACK if the Gateway target MCP server is unreachable. Error: Failed to connect and fetch tools from the provided MCP target server.
agentcore remove all --force only resets local config files. AWS resources remain until you run agentcore deploy to apply the empty state.
agentcore invoke --session-id with a short string fails with Member must have length greater than or equal to 33. Use UUIDs or long identifiers.
AgentCore Runtime's ENI in the VPC may stay in-use for hours after agentcore destroy. Wait for it to become available before deleting the subnet and VPC.
Mixing up bedrock-agentcore-control (create/delete) and bedrock-agentcore (session ops/code execution) gives you an AttributeError.
Uploading a raw Python file to S3 for create-agent-runtime gives 'S3 object must be in .zip format' ValidationException. You must zip the code first.
The docs say 'Stateless between commands' but the filesystem is actually shared within a session. Only shell history and environment variables are reset.
Setting entryPoint to ['python3', 'main.py'] in codeConfiguration gives 'Invalid entrypoint value' error. For Python runtimes, pass only the filename with .py extension like ['main.py'].
agentcore deploy emits 'X-Ray Delivery Destination is supported with CloudWatch Logs as a Trace Segment Destination' during memory creation, but the runtime phase auto-configures it.
agentcore configure with Direct Code Deploy fails with 'Direct Code Deploy deployment unavailable (uv not found)'. Fix with pip install uv — not mentioned in the official docs.
Creating a Cedar permit without a when clause in Bedrock AgentCore Policy fails with CREATE_FAILED due to FAIL_ON_ANY_FINDINGS automated reasoning.
Bedrock AgentCore Gateways only allow hyphens while Policy Engines only allow underscores. Using the same prefix for both causes ValidationException.
batch-update-memory-records requires timestamp, and memory-strategies uses semanticMemoryStrategy / namespaceTemplates — not what the docs suggest.