Posts
All articles on AI-native work patterns, agent infrastructure, and what actually works when running agents as real tools.
-
Does Graphify Actually Help an AI Coding Agent?
A code graph earns trust only when it beats disciplined source search on a correct, source-backed task.
-
Why AI Coding Tools Are Getting Cheaper: Prompt Caching Explained
Updated:Prompt caching discounts repeated input, but the savings depend on exact prefix reuse, model-specific retention, and enough reads to repay each cache write.
-
Prompt Caching Is a Harness Design Problem, Not Just an API Toggle
A 186-request test shows why cache savings depend on stable tools, deterministic prefixes, deliberate breakpoints, routing keys, and measurement inside the agent harness.
-
Building AI Voice Agents with Amazon Nova Sonic and LiveKit: Patterns, Architecture, and a RAG Latency Surprise
Patterns that held up across a series of Nova 2 Sonic + LiveKit voice agents, a telephony architecture that took a real phone call, and a latency finding that contradicts the standard advice on voice RAG.
-
FLUX.1 LoRA fine-tuning on fal.ai is remarkably simple
A tested workflow for training a FLUX.1 style LoRA on fal.ai, applying it through the current API, and understanding where consistency still falls short.
-
Web Search as a Managed Connector: Wiring Amazon Bedrock AgentCore WebSearch Into Any Agent
Amazon Bedrock AgentCore WebSearch is a managed MCP connector, not a search-API wrapper: an Amazon-operated index with zero data egress, IAM-authorized so there are no secrets to rotate, and $7 per 1,000 queries confirmed on the bill.
-
I Pointed an AI Code-Security Tool at My Own Control Plane
Corridor reviews AI-generated code inside the coding loop instead of after commit. I imported a real AWS control plane, planted an IDOR bug in a PR, and watched what it caught.
-
What Actually Gets Stored When You Put a Vector in S3 Vectors
A vector store keeps three things per record — a key, an array of floats, and a bag of metadata. Your document text is stored as metadata, and a 2 KB limit on the filterable part is what quietly breaks ingestion.