Read what ships.

01ByteByteGo · architecture

A Deep Dive into Salesforce’s Agent Tech Stack

Salesforce transitioned its AI agent platform from a single, complex prompt architecture to a multi-layer design combining deterministic code with reasoning models to overcome instruction loss. The new system delegates rule enforcement, data harmonization, context management, and permissions to dedicated backend components. This approach significantly reduces token consumption and improves operational reliability in production environments.

DEPTH ■■■■■13 min readyesterday

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 read11 days ago

Machine Learning

Sign in to save and rate

03ByteByteGo · migration

Why Shopify Migrated From React Native

Shopify migrated its mobile applications from React Native to native Swift and Kotlin because AI coding agents made the cost of duplicating features across platforms negligible. Using a greenfield approach and a custom tool called Helix to break down the migration into small, reviewable steps, the team successfully rewrote the apps. The resulting native applications showed significant improvements in startup time, crash rates, and app size.

DEPTH ■■■■■11 min read7 days ago

Sign in to save and rate

04Kubernetes · migration

The Shift to cgroup v2 in Kubernetes: What You Need to Know

Kubernetes has deprecated cgroup v1 in favor of cgroup v2, with complete removal scheduled for future releases. The transition requires migrating Linux nodes, updating container runtimes, and verifying kernel compatibility. Adopting cgroup v2 enables advanced features such as Memory QoS, improved OOM handling, and pressure stall information metrics.

DEPTH ■■■■■12 min read3 weeks ago

Sign in to save and rate

05ByteByteGo · architecture

How Uber Built a Genie to Answer Questions Correctly

Uber built an internal AI assistant called Genie to answer engineering support questions directly within Slack. The system uses a retrieval-augmented generation (RAG) architecture powered by Apache Spark, LangChain, and vector databases to fetch and explain documentation. As a result, Genie reduces repetitive on-call support burdens while providing source attribution and fallback paths to human engineers.

DEPTH ■■■■■13 min read8 days ago

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

07Kubernetes · scaling

Scaling Kubernetes Workloads with Node Swap

Kubernetes memory limits often cap node density for bursty agentic AI workloads that sit idle with large memory footprints. To solve this, the engineering team benchmarked Kubernetes v1.34 node swap backed by fast NVMe Local SSDs across CI/CD builds, headless browsers, and Python sandboxes. The approach successfully increased pod density up to three times without significant latency costs.

DEPTH ■■■■■6 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

09ByteByteGo · architecture

LLM Prompt Caching Versus Semantic Caching

This post compares prompt caching and semantic caching for large language models to reduce costs and latency. Prompt caching reuses intermediate KV states for shared prefixes, while semantic caching stores previously generated answers for similar questions using vector embeddings. The article explains the mechanics, pricing models, and trade-offs of both approaches.

DEPTH ■■■■■16 min read6 days ago

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

11Stack 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

12ByteByteGo · architecture

How Netflix Taught an LLM to Recommend Movies So That You Keep Watching

Netflix developed GenRec, an LLM-based recommendation ranker designed to interpret viewing histories and item metadata in plain text to score movies and TV shows. By adapting an open-source foundation model in a two-phase training process and employing context engineering to reduce token usage, the team improved catalog ranking and satisfaction prediction. The system balances general language knowledge with domain-specific constraints to manage costs and personalization.

DEPTH ■■■■■15 min read2 weeks ago

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

14ByteByteGo · architecture

Why LLMs Agree With You Even When You’re Wrong

Large language models often agree with incorrect user claims because their training rewards human approval and helpfulness over strict accuracy. This behavior, known as sycophancy, is driven by reinforcement learning from human feedback and conversational pressure. Teams can mitigate this issue by using synthetic training data, Constitutional AI principles, and internal activation probes to penalize agreeable falsehoods.

DEPTH ■■■■■13 min read3 weeks ago

Sign in to save and rate

15Kubernetes · tooling

Why Is My Kubernetes Job Still Pending? Tracing Kueue Admission with Headlamp

Kubernetes jobs managed by Kueue can remain pending while waiting for admission checks and resource allocation. The team implemented a Kueue plugin for Headlamp to trace pending workloads through local queues, cluster queues, and resource flavors within a unified interface. This approach allows engineers to quickly determine whether a job is blocked by admission policies or standard pod scheduling.

DEPTH ■■■■■9 min read13 days ago

Sign in to save and rate

16Stack 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

17ByteByteGo · performance

The LLM Blindspot: Why Models Forget What’s in the Middle of Your Prompt

Large language models frequently fail to utilize relevant information when it is placed in the middle of a long prompt due to architectural attention biases. To mitigate this issue, engineering teams can optimize prompt organization, remove unnecessary context, and apply effective retrieval strategies. Ultimately, expanding context capacity does not automatically guarantee reliable information retrieval across all positions.

DEPTH ■■■■■11 min read3 weeks ago

Sign in to save and rate

18ByteByteGo · architecture

How to Build Models Without Language, with Logical Intelligence

This article explores the limitations of large language models on constraint satisfaction problems like Sudoku and introduces energy-based reasoning models as an alternative. It explains how energy-based models utilize continuous latent spaces and descending energy gradients rather than left-to-right token generation. The result is a more reliable approach for non-language reasoning tasks.

DEPTH ■■■■■19 min read2 weeks ago

Sign in to save and rate

19Stack 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 read4 weeks ago

Machine Learning

Sign in to save and rate

20Stripe · practice

Why I tried to kill token billing (and why we kept it)

The author examines the limitations of token-based billing as a customer-facing pricing model, noting that it exposes internal margins and commoditizes AI products. The company explored abandoning token billing entirely but realized its necessity for backend cost tracking and margin management. Ultimately, the piece proposes unified credits and output-based pricing as better ways to align invoices with customer value.

DEPTH ■■■■■7 min read3 weeks ago

Sign in to save and rate