Explore fresh perspectives and expert takes on design, development, and digital innovation.
BM25 + vector + RRF is often treated as a default “better retrieval” baseline, but it can flatline or regress once you apply real filters, real latency budgets, and real query mix. This article shows how to design an apples‑to‑apples retrieval bake‑off that exposes those failure modes and produces a decision you can ship.
Read more →
EMA makes your IdP the source of truth for which MCP tools exist and who can use them. That shifts rollouts from “each app does OAuth” to centrally governed provisioning, with different failure modes you need to design for.
Read more →
Teams keep discovering that workflow engines excel at retries, timers and orchestration, but ‘resume’ for non-deterministic agent loops is a different problem. A practical 3-layer pattern (loop, durability substrate, workflow orchestrator) keeps runs debuggable without exploding workflow history.
Read more →
MCP tool sprawl turns context into a tax you pay on every turn. This piece lays out an evaluation harness to compare always-inject schemas, tool search (deferred loading), and grouped progressive discovery using token footprint, tool-pick accuracy, and end-to-end latency.
Read more →
Large MCP servers are turning flat tool lists into context bloat, usability failures, and brittle access control. Teams are converging on three patterns — primitive grouping, client-side tool search/skills, and progressive discovery — and each breaks in different ways.
Read more →
Agentic automation fails in production for a boring reason: actions happen, but nobody can reliably prove which principal authorised each tool call, under which policy, with which inputs. This ‘agent identity gap’ turns good design docs into weak operational security.
Read more →
MCP’s 2026-07-28 spec pushes remote tool servers into a stateless, HTTP-native shape. That changes where you terminate identity, how you gate and version tools, and how you run them behind the same controls as any other API.
Read more →
Agents rarely fail because your prompt is slightly off. They fail because “resume after tool failure” is underspecified: retries, partial progress, and state handoffs aren’t durable, so you can’t safely continue — and you can’t observe what really happened.
Read more →
Hybrid (BM25 + vectors) plus a cross‑encoder reranker is now the default RAG advice, but it can make real systems worse. Here’s a practical, eval-driven way to decide when to rerank using similarity margins and failure-mode buckets.
Read more →