Elizabeth Fuentes L

Everyone is racing to make agents remember more. Bigger context windows, longer histories, a vector...

📦 Clona y dale ⭐ a stop-ai-agents-losing-memory-sample-for-aws Los agentes de IA acumulan hechos...

📦 Clone and ⭐ stop-ai-agents-losing-memory-sample-for-aws AI agents accumulate facts across...

AI agent observability means capturing your agent's reasoning cycles, tool calls, and token usage as...

Amazon DynamoDB Vector Search GA: agrega un índice vectorial a una tabla existente. Misma precisión que S3 Vectors, latencia de un solo dígito en ms, vectores y datos operacionales en el mismo lugar.

Amazon DynamoDB Vector Search GA: add a vector index to an existing table. Same accuracy as S3 Vectors, single-digit ms latency, data and embeddings collocated.

La búsqueda por palabras clave pierde preguntas de intención. La búsqueda semántica las encuentra con FAISS o Amazon S3 Vectors, medido.

Keyword search misses intent questions. Semantic search finds them. Add vector search to AI agent memory with FAISS or S3 Vectors, benchmarked.

Detecta alucinaciones en agentes de IA sin datos etiquetados. Detección LSC zero-shot,...

Claude Fable 5 falla en Amazon Bedrock con un error 400 antes de procesar un solo token: "data...

Tu agente de IA funciona perfecto en el demo. Luego un usuario real regresa al día siguiente y el...

Here's a test most AI agents fail. A brand-new user searches flights, books one in business class, and asks what you recommend based on what you know about me?

What Vibe Coding Skips (And Why Specs Caught It First) ...

Your AI agent works beautifully in the demo. Then a real user comes back the next day, and the agent...

Integrating Lambda Durable Functions into a Step Functions Workflow ...

Un pipeline RAG fintech envenenó su propio vector store y el validador LLM-as-a-judge aprobó cada alucinación. La solución: validar las escrituras con código.

💻 Todo el código de esta serie está en un solo repo: resilient-agent-harness-sample-for-aws. Este...

💻 Todo el código de esta serie está en un solo repo: resilient-agent-harness-sample-for-aws. Este...

A fintech RAG pipeline poisoned its own vector store and the LLM-as-a-judge validator approved every hallucination. The fix: gate writes with code.

💻 Todo el código de esta serie está en un solo repo: resilient-agent-harness-sample-for-aws. Este...

💻 Todo el código de esta serie está en un solo repo: resilient-agent-harness-sample-for-aws. Este...

💻 Este es el inicio de una serie. Todo el código está en un solo repo:...

An AI agent that's flawless in the demo can still fall apart the first time a tool fails in production: a timeout, a network error, a response that comes back corrupted. The fix isn't a smarter model. It's testing the agent against those failures on purpose, before your users do, and hardening it one failure type at a time.

A static AI agent re-reasons the same kind of task from scratch every time, burning tokens and sometimes getting it wrong differently on each run. A self-improving agent solves it once, writes a small verified tool, and reuses that tool exactly from then on. The catch worth knowing up front: writing the tool costs more tokens than one-off reasoning. The payoff is correctness and reuse, not a smaller bill.

On a multi-step task, an AI agent will trust a tool that reports success even when the work silently never saved, and then confidently report the whole task done while part of it quietly failed. The fix isn't a better prompt. It's checking each step against the real backend and re-doing the one that didn't take.

When an AI agent reads untrusted content (a web page, a document, an email), a hidden instruction can ride in, get stored in the agent's own memory, and fire in a later session to exfiltrate data. This persistent form of prompt injection is called memory poisoning, and a better prompt won't stop it. The reliable defense is a deterministic gate on the dangerous action, at the tool boundary.

When an AI agent hallucinates a fact, the real damage starts when it writes that fact to memory and re-reads it as trusted context every session after, compounding the error and the token cost. Stopping it isn't a prompting problem. It's a deterministic check that validates what the agent is about to write, so a bad fact never becomes a permanent memory.

De cero a la nube: cómo dockericé mi feed de AWS y lo desplegué en ECS ...

Un agente de IA tiene dos tipos de memoria: conversación (semántica) y contexto (referencia exacta). Mantenlas separadas con Strands y AgentCore.

An AI agent has two kinds of memory: conversation (semantic) and context (exact reference). Keep them separate with Strands and AgentCore.