Samuel Tambunan's Blog

Weekly .NET & AI Briefing — September 27–October 3, 2026

This week’s reading list leans heavily on one theme: AI agents are now a first-class architectural concern, not a side project. Microsoft shipped agentic UI components for Blazor, Uber built a company-wide gateway just to govern agent access to its services, and Milan Jovanović dissected GitHub’s own agent-assisted Rust rewrite of the Copilot runtime. Alongside that, there’s solid distributed-systems and security material from Netflix, Elastic, and Airbnb worth your time if you’re building toward a software-architect or engineering-leadership role.

.NET, ASP.NET Core & Architecture

  1. What It Actually Took to Port the GitHub Copilot Runtime to Rust — Milan Jovanović’s Blog, October 3 GitHub rewrote the Copilot agent runtime from TypeScript/Node.js into over 800,000 lines of Rust so it could be loaded in-process across six SDK languages via a C ABI, instead of paying separate-process overhead. The 14.5-week port leaned on coding agents for most of the code generation while engineers focused on architecture, tests, and system integrity — Milan’s framing, “the agents changed the amount of code one engineer could supervise,” is a useful line for thinking about how agentic tooling reshapes engineering-team structure, not just individual productivity.

  2. Build Agentic UI with the new Blazor AI components — .NET Blog, September 28 New experimental Blazor components turn agent interactions (streamed output, tool calls, human-approval steps, shared state) into renderable, controllable content blocks, built on Microsoft.Extensions.AI, AG-UI, and the Microsoft Agent Framework. Worth a look if you’re designing UI patterns for human-in-the-loop agent workflows rather than plain chatbots.

  3. Aspire 13.6: Your dashboard gets memory — .NET Aspire Blog, September 29 Aspire’s dashboard now persists traces and logs across restarts so you can compare runs instead of losing telemetry every time you relaunch the stack, plus in-dashboard container shell access and first-party Java and Rust hosting. Relevant for architects standardizing local/dev-loop tooling across polyglot services.

Agentic AI in Production

  1. Designing MCP Gateway: Uber’s MCP Management Platform — Uber Engineering, October 1 Rather than let every team bolt together its own Model Context Protocol integration, Uber built a centralized gateway that translates MCP calls into HTTP/gRPC/TChannel and auto-discovers existing internal APIs as agent-usable “tools.” It now fronts 800+ MCP servers and 5,000+ tools company-wide — a concrete case study in governing agent-to-service access at scale instead of letting it sprawl.

  2. The best LLM writes correct Elasticsearch ES|QL 59% of the time — Elastic Search Labs, September 28 Elastic ran 6,000 LLM-generated ES|QL queries from four models against 500 natural-language questions and found the strongest model got it right on the first try only 59% of the time, failing mostly on join-key mismatches, SQL-syntax habits bleeding into ES|QL, and models guessing at values not in the dataset. A sobering, data-backed reminder that “grammar is no longer the ceiling” for text-to-query systems — data familiarity is.

  3. Personalization without user identity — The Airbnb Tech Blog, September 29 To solve cold-start personalization without relying on persistent user identifiers, Airbnb clusters nearby users (via coarse IP geolocation) into groups of roughly 1,000 people and feeds aggregated behavioral signals into ML models through a “proximity key” — any model that currently conditions on a user ID can swap in the proximity key instead. A nice architectural pattern for privacy-constrained personalization.

Big-Tech Engineering & Infrastructure

  1. Netflix Stratum Media Processing: Automated Container Right-Sizing — Netflix Technology Blog, September 25 Netflix describes how its media-processing platform automatically right-sizes containers instead of relying on static, hand-tuned resource requests — a practical look at closing the loop between workload profiling and infrastructure allocation at Netflix’s scale.

  2. Trading a Cloud Identity for Your Own: Workload Attestation on Managed Compute — Netflix Technology Blog, September 25 Netflix bridges cloud-provider identity and internal application identity through a three-way check: the control plane signs a claim about the workload, the compute driver proves possession of its cloud credentials, and the identity service corroborates both before issuing a certificate — because “neither statement is sufficient by itself.” A clean case study in zero-trust identity design for managed compute.

  3. Check 100 candidates, not 10 million documents: Faster kNN filters in Elasticsearch — Elastic Search Labs, October 2 Elasticsearch now automatically chooses between pre- and post-filtering per kNN query, evaluating filters only against candidate vectors rather than whole segments; post-filtering won in 104 of 120 benchmarked cases on a 10M-vector corpus while a binomial model sizes the candidate oversampling to guarantee correct results.