Craft & Leadership

Architecture, Practice & Running Teams. Published every Saturday, here and by RSS.

38 stories 7 issues 9 categories

Architecture Leading Teams Incidents & Ops Testing & Quality Migrations Productivity Careers Opinion Practice

Top stories

DoorDash Uses Multi Agent LLMs to Clean up 60,000 Feature Flags

DoorDash designed a multi-agent LLM system using Model Context Protocol (MCP), parallel agents, and isolated Git worktrees to automate the removal of stale feature flags. The system successfully resolved 45 out of 50 evaluated flags, generating usable pull requests in under 14 minutes at an average cost of $4.79 per cleanup.

Why it matters — Automating tedious codebase maintenance with LLMs can be highly cost-effective ($4.79 per cleanup) when combined with automated validation and isolated worktrees to ensure safety.

InfoQ Architecture · 19 September 2026 · Read the original →

Migrations

Agoda Replaces 72-Shard SQL Server Price Cache with DragonflyDB

Agoda migrated its 1.5 TB hotel price cache from a complex 72-shard SQL Server setup to DragonflyDB to handle scaling read/write demands. The migration was executed safely using dual reads, parity validation, and gradual traffic shifting, resulting in an eightfold reduction in P99 read latency.

Why it matters — Shifting a high-throughput cache from relational shards to an in-memory store can yield an 8x reduction in P99 latency when managed via a staged dual-read and parity validation strategy.

InfoQ Architecture · 19 September 2026 · Read the original →

Testing & Quality

How well do agents use test and verification techniques?

This post investigates whether instructing AI coding agents to use specific testing and verification libraries actually improves the correctness of their code implementations. It explores the gap between the ease of hitting quality bars with agents and the observed decline in overall software quality.

Why it matters — Instructing AI agents to use specific test and verification techniques is a critical lever for improving the correctness of generated code and reversing the trend of declining software quality.

Dan Luu · 19 September 2026 · Read the original →

Leading Teams

Nobody needs a human router

This post examines the evolution of the engineering manager role, arguing that the traditional focus on acting as an information router—translating context up and down the hierarchy—is no longer sufficient. It highlights how the role must shift away from merely smoothing communication friction.

Why it matters — Engineering managers must transition from being information routers to driving higher-value organizational impact, as simply translating context up and down is an outdated optimization.

The Engineering Manager · 19 September 2026 · Read the original →

Opinion

Fragments: September 8

This post highlights an asymmetric challenge in the age of AI: while the cost of generating code and content has plummeted, the cost of verifying its correctness remains high. This explains why early AI successes are concentrated in areas where outputs are easy for humans to inspect.

Why it matters — The core bottleneck of AI-assisted engineering is not generation cost, but verification cost; tools succeed most where outputs are cheap and easy for humans to inspect.

Martin Fowler · 19 September 2026 · Read the original →

Browse all 38 stories in this channel →33 older stories

We are finally automating the cleanup of our own legacy decisions.

Weekly editions

Each edition is one week's digest exactly as it was published — a summary of the week, then the stories.