All Posts

Every agent I run had perfect amnesia between messages. Wiring Qdrant Cloud into the fleet fixed that — and the four undocumented constraints between a running vector database and a working one turned out to be the whole story.

A Telegram message asking which bin to put out becomes a three-agent A2A delegation chain — and 278 OpenTelemetry spans. Here's how to see the whole conversation in Langfuse and cut 79% of the noise with a filter processor.

Why a codebase knowledge graph beats grep for agents, how to run one on Kubernetes without building a container image, and the six failures between a working binary and a working service.

Anatomy of a CRITICAL under-replicated partitions alert on self-managed Kafka running on EKS, and why Kubernetes Reflector is the unglamorous tool that kept MirrorMaker2 from breaking during a CA rotation.

How the kagent KMCP controller makes MCP servers a normal workload on EKS, why the caller should carry the credential instead of the server, and what connecting an organization to MCP does for delivery time.

Installing Terraform Enterprise FDO on EKS behind ArgoCD and cdk8s looks straightforward until TLS certs, Traefik, and SAML bootstrap each throw a silent failure with no obvious cause.

Why an AI gateway isn't optional infrastructure — and what happens to platforms that skip it.

How to wire up a Bedrock Agent that uses Lambda-backed action groups as MCP tools — PagerDuty, Jira, Confluence, and Slack — attached to the Knowledge Base from the previous post.

A field guide to provisioning a Bedrock Knowledge Base on Aurora pgvector with Terraform, then keeping it in sync with a GitHub Actions workflow that ships docs to S3 on every push.

How we built sys-tag-manager — a Go service using Clean Architecture and AWS Resource Explorer — to automate cloud tagging, eliminate orphaned resources, and give finance teams accurate cost attribution.

EKS 1.32 replaces the legacy bootstrap.sh script with nodeadm. Here's what changed, why it matters, and how to migrate before the AL2 end-of-support deadline.