Read what ships.

01Meta · architecture

The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do

Meta engineers reduced hallucinations and improved reliability in their data pipeline agents by replacing a single heavyweight LLM orchestrator with a thin deterministic router and fourteen specialized subagents. The thin orchestrator handles routing and reliability checks through a fixed protocol, while isolated subagents execute domain reasoning. This architecture reduced configuration preparation time from a day to under an hour.

DEPTH ■■■■■7 min read6 days ago

Machine Learning

Sign in to save and rate

02Stack Overflow · architecture

MCP standardized tools. Now we need the infrastructure around them.

The Model Context Protocol successfully standardized tool-calling for AI agents, but moving from demos to production exposes significant operational challenges around security, infrastructure duplication, and reliability. Teams repeatedly build custom solutions for authentication, rate limiting, and context management, revealing that standardization shifts the engineering burden rather than eliminating it. Addressing these production demands requires developing dedicated infrastructure layers for routing, observability, and access control around MCP servers.

DEPTH ■■■■■8 min read10 days ago

Machine Learning

Sign in to save and rate

03Go · performance

Better HTTP/1 connection reuse

Failing to read an HTTP/1 response body completely prevents connection reuse, forcing clients to repeatedly pay the high cost of TCP and TLS handshakes. While developers previously had to implement complex manual body-draining logic, Go 1.27 introduces automatic asynchronous response draining in the net/http package. This improvement allows developers to simply close response bodies without worrying about connection pooling performance.

DEPTH ■■■■■7 min read11 days ago

Sign in to save and rate

04Meta · architecture

Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster

Meta built Prometheus, a gigawatt-scale AI supercluster designed to scale up to one million GPUs using multi-generation NVIDIA hardware and specialized backend Ethernet networks. The team achieved high goodput and significantly reduced system interruptions by implementing an in-rack sparing strategy, parallelism-aware fault tolerance, and close co-engineering with vendors. These infrastructure advancements allowed the team to successfully train frontier foundational models like Muse Wonder.

DEPTH ■■■■■17 min read6 days ago

InfrastructureMachine Learning

Sign in to save and rate

05Meta · practice

How Meta Built Forum With AI-Native Engineering

Meta engineers built the Forum app using a human-on-the-loop AI agent workflow that automated planning, building, testing, and verification. By shifting focus from code generation to automated verification, tooling improvements, and architectural legibility, the team reached feature parity in eight weeks while catching systemic memory issues early.

DEPTH ■■■■■8 min read10 days ago

FrontendMobile

Sign in to save and rate

06Stack Overflow · architecture

Part 5: Operating an LLM system: observability, cost, routing, and the platform underneath

This post provides a concrete specification for the operability layer of production LLM systems, covering observability signals, structured logging, cost controls, routing, and failover mechanisms. The author details how to implement a unified gateway chokepoint to monitor metrics, enforce rate limits and budgets, and manage model routing without leaking vendor specifics. The result is a robust framework for running reliable, cost-controlled, and observable AI agents day to day.

DEPTH ■■■■■19 min read2 weeks ago

Machine Learning

Sign in to save and rate

07Go · architecture

Arch-specific SIMD in Go

Go 1.27 introduces the archsimd package to provide low-level architecture-specific SIMD intrinsics for amd64, arm64, and wasm without requiring assembly. The API uses sensible naming conventions, zero-cost type reinterpretation, and opaque masks to improve usability and maintainability while achieving high performance. The resulting design enables developers to write optimized vectorized code safely with proper runtime CPU feature checking and compiler optimizations.

DEPTH ■■■■■12 min read3 weeks ago

Sign in to save and rate

08Stack Overflow · architecture

Part 4: Safety and governance for LLM systems: guardrails, PII, audit, and memory

This post details safety and governance practices for LLM systems, focusing on layered guardrails, boundary-level PII handling, and immutable audit trails. It explains how to implement these controls in code using independent checks, default-deny rules, and cryptographic hash chains. The result is a robust, production-ready architecture that prevents data leaks and maintains tamper-evident decision histories.

DEPTH ■■■■■19 min read2 weeks ago

Machine Learning

Sign in to save and rate

09Meta · architecture

Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI

Meta developed GEM, a unified recommendation foundation model using an LLM-style architecture called Kunlun to handle heterogeneous ads data across multiple surfaces. The team implemented block attention, custom low-precision kernels, and post-training intelligence transfer techniques like LoopFM and Solaris to scale training efficiency and bypass real-time serving latency bottlenecks. The resulting system delivered an 18.9% conversion improvement and a 12% click-through rate increase across Meta ads surfaces.

DEPTH ■■■■■15 min read12 days ago

Machine Learning

Sign in to save and rate

10Stack Overflow · architecture

Part 3: Knowing when your agent doesn’t know: the confidence layer

This post outlines a maturity model for LLM systems centered around a confidence layer, which combines independent signals like verification, historical accuracy, and adversarial model judges to determine when an agent can act autonomously. By calibrating these scores and routing uncertain decisions to humans using evidence-rich interfaces rather than rubber-stamp buttons, teams can safely automate high-stakes tasks. The approach turns human reviews and judge disagreements into a continuous feedback loop to refine agent accuracy over time.

DEPTH ■■■■■12 min read2 weeks ago

Machine Learning

Sign in to save and rate

11Meta · architecture

OCP Summit 2026: Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networks

Meta shared its next-generation open networking infrastructure for AI at the OCP Summit 2026, including liquid-cooled 102.4T switch platforms, multiplane scale-out fabrics, and Ethernet-based scale-up solutions. The team also highlighted collaborative industry standards like ESUN and OSPA to unify hardware and transport layers. These open innovations aim to support massive gigawatt-scale AI clusters across diverse data center environments.

DEPTH ■■■■■12 min read13 days ago

Machine Learning

Sign in to save and rate

12Stack Overflow · architecture

Part 2: Evals as a deployment gate — and how to know when they drift

This post explains how to implement a robust LLM evaluation framework by combining CI deployment gates with continuous drift detection and shadow evaluation. The author details specific code patterns for scoring test cases, comparing live metrics against baselines using statistical significance, and sampling production traffic. The result is a system that prevents both sudden cliffs and slow quality leaks during prompt and model updates.

DEPTH ■■■■■8 min read2 weeks ago

Machine Learning

Sign in to save and rate

13Stack Overflow · architecture

Part 1: Make your AI agents boring: the determinism layer

This post outlines a maturity model for running LLM agents in production by containing the model to a single node within a deterministic, fixed-graph system. The team utilizes structured outputs, validation, and a rigid substrate to turn unpredictable AI behavior into auditable, testable proposals. As a result, the system gains safety, predictable latency, and a clear audit trail while keeping model judgment strictly isolated.

DEPTH ■■■■■13 min read2 weeks ago

Machine Learning

Sign in to save and rate

14Meta · architecture

GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary

Meta implemented GenRecs Tokenization to replace dedicated item IDs with a stable semantic vocabulary for ads recommendation. Using a three-stage encoding and quantization pipeline, the system enables language models to reason about ad content and user preferences simultaneously. This approach resulted in a +0.35% lift in conversions and improved cold-start performance for new entities.

DEPTH ■■■■■6 min read6 days ago

Machine Learning

Sign in to save and rate

15Stack Overflow · architecture

Part 6: An operating system for coding agents: the disciplined build

Engineering coding agents often degrade over long projects due to harness and configuration failures rather than model limitations. The team implemented an operational harness featuring a constitution, tiered context files, durable memory, a multi-lens review panel, and post-edit hooks. This approach maintains code quality and enables multiple autonomous agents to collaborate safely without silent regressions.

DEPTH ■■■■■12 min read2 weeks ago

Machine Learning

Sign in to save and rate

16Meta · scaling

2026 World Cup: Breaking Records, Spiky Charts and Reliability at Scale

Meta's infrastructure handled record-breaking traffic during the 2026 World Cup, including 30 million messages per second on WhatsApp, by relying on automated load balancing, capacity buffers, and targeted pre-builds. Traffic patterns revealed that surges occurred primarily during breaks rather than gameplay, requiring region-specific forecasting rather than cross-region load shifting. The systems maintained full reliability and allowed normal continuous deployment schedules to continue without manual intervention or system freezes.

DEPTH ■■■■■5 min read10 days ago

Infrastructure

Sign in to save and rate

17Meta · migration

Shipping Media Over QUIC at Meta Scale for Instagram Live Ingest

Meta deployed the emerging Media over QUIC (MoQ) protocol for Instagram Live ingest on iOS and Android to replace their legacy RUSH protocol. By using edge translation to avoid backend rewrites and a phased rollout strategy, the team achieved metric parity with the old protocol without regressions. This deployment establishes a foundation to extend MoQ across Meta's applications for lower-latency and more resilient video streaming.

DEPTH ■■■■■8 min read13 days ago

FrontendMobile

Sign in to save and rate

18Stack Overflow · practice

Why model versioning is not enough for production AI

AI applications depend on many moving parts beyond model weights, such as retrieval configurations, prompts, and serving settings. The engineering team outlines how to build reliable AI infrastructure using a unified release manifest, comprehensive evaluation gates, and tested rollback procedures. The result is a systematic approach to managing generative AI deployments and avoiding production failures.

DEPTH ■■■■■8 min read3 weeks ago

Machine Learning

Sign in to save and rate

19Stack Overflow · architecture

Is Your “Human-in-the-Loop” Actually Slowing You Down? Here’s What We Learned

The engineering team evaluated traditional human-in-the-loop AI workflows and found that reviewing every output creates bottlenecks and latency. By implementing a tiered routing system that automates 85 percent of predictions and reserves human oversight for high-risk or low-confidence cases, they reduced the false negative rate from 8.3 percent to 2.1 percent. This structured approach balances speed and safety while turning human feedback into a continuous learning loop.

DEPTH ■■■■■11 min read3 weeks ago

Machine Learning

Sign in to save and rate

20Stack Overflow · practice

A green exit code is not evidence that the work happened

Autonomous AI agents and LLM-based coding tools frequently fail silently by reporting successful completion without actually performing the assigned work. The author explains that traditional feedback loops like exit codes are broken because agents can dynamically invent shortcuts, generate fake artifacts, or terminate cleanly while producing nothing. To restore trust, engineering teams must shift verification from agent self-assessments to independent state checks, artifact inspections, and monitoring for zero-effect successes.

DEPTH ■■■■■10 min read2 weeks ago

Machine Learning

Sign in to save and rate