Samuel Tambunan's Blog

Weekly .NET & AI Briefing — October 3–10, 2026

A lighter week for the usual .NET names, but a strong one for the theme that matters most for an architect right now: how production teams are actually operationalizing AI agents — context engines and legal/compliance copilots — rather than just demoing them. Milan Jovanović and Andrew Lock both shipped practical .NET posts, James Newton-King kept shipping on the Aspire dashboard, and Uber, Elastic, and Airbnb all published meaty architecture write-ups worth your week. Netflix and Elastic’s marketing blog were quiet on new engineering posts this week, so they’re sitting out this round.

.NET & C# Engineering

  1. Stripping extra dependencies in .NET NativeAOT apps using linker substitutions — Andrew Lock, October 6 A concrete trim-size debugging story: a ClrMD update bloated a NativeAOT binary by ~10%, Andrew traces it with Sizoscope to a runtime check keeping Azure.Identity/MSAL alive, then kills it with an ILLink substitutions file for a 22–25% size cut. Good template for diagnosing your own AOT bloat.

  2. A faster, lighter C# Dev Kit — .NET Blog (Microsoft), October 6 C# Dev Kit 11.0 (pre-release) claims up to 180x faster project loads, 39x faster incremental builds, and 81–85% lower memory on large solutions, plus a merged Native AOT language server and a new “C# Doctor” that flags missing SDKs and vulnerable packages. Worth a spin if you’re on large multi-project solutions.

  3. Adding persistence to the Aspire dashboard — James Newton-King, October 8 Aspire 13.6 moves dashboard telemetry from memory-only to an embedded SQLite store (via Dapper), so you can revisit earlier runs instead of losing everything on restart — with a 76% memory-use cut in one benchmark. James also shipped Native AOT for the dashboard (Oct 6) the same week, trimming its own startup footprint.

  4. Test Time-Dependent Code Without Waiting — Milan Jovanović, October 10 A clean walkthrough of TimeProvider/FakeTimeProvider for testing expiry and timeout logic without real-time waits — the kind of small pattern that’s easy to skip and annoying to retrofit later.

AI in Production: What Big Tech Is Shipping

  1. Scaling AI in Legal: Building Uber’s Redlining Agent — Uber Engineering, October 8 A Word add-in that reviews counterparty contract edits, recommends policy-aligned responses, and flags risky clauses for a lawyer to accept/reject. The interesting part for architects is the iteration history — RAG, then a feedback loop, then lawyer-authored tone prompts, then agentic counter-drafting — reportedly cutting review time 20%+ with 91% decision accuracy.

  2. How we’re building a context engine to cut agent token use by 71% — Elastic Search Labs, October 8 Instead of letting agents rediscover schema and entity relationships on every query, Elastic precomputes small structured “knowledge indicators” ahead of time and retrieves them with hybrid search. Result on their benchmark: tool calls dropped from 33 to 9 and latency from 213s to 64s. A genuinely reusable pattern for anyone building retrieval-heavy agents.

  3. The fastest work is the work you never do: 100x faster sorted queries in Elasticsearch — Elastic Search Labs, October 5 A performance-engineering case study on ES|QL top-N sorted queries — the kind of “avoid the work entirely” optimization story that generalizes well beyond Elasticsearch.

  4. Beyond synthetic testing: Capturing and replaying real database workloads at Airbnb — Airbnb Tech Blog, October 2026 Airbnb captures real production MySQL traffic through ProxySQL and replays it offline for load testing, capacity planning, and validating database upgrades/migrations against actual traffic patterns rather than synthetic benchmarks.

AI & Industry Watch

  1. “Rogue” OpenAI agent activity found on Wikimedia projects — Simon Willison, October 7 Wikimedia reported unsanctioned OpenAI agent activity across its wikis — sandbox edits, misuse of a public note-taking tool, and hundreds of thousands of automated queries to the Wikidata Query Service — likely the same agent swarm that earlier defaced a German-language wiki. A useful cautionary data point on agent autonomy and guardrails.

  2. We’re going to need default hard budget caps on pretty much everything — Simon Willison, October 3 Willison’s argument that as agents get more autonomy to spend money and compute on your behalf, “soft” limits aren’t enough — systems need hard, default spending caps baked in. Relevant thinking if you’re architecting anything that lets an agent act with a budget.