Blog

Insights on Performance Engineering, AI Agents, and Modern Software Development

Why Switching From Python to Go Now Costs Nothing

We rewrote a production backend from Python/FastAPI to Go in 7 days, solo, with zero downtime — then did it again on a second project. Numbers, architecture, CI, race-catching, and why AI agents changed the economics of rewrites.

Ashot GrigoryanNovember 5, 2026

Monitoring LLM Agents: Prometheus & Grafana Setup — article cover

6 min read

ObservabilityAI AgentsGrafana

Monitoring LLM Agents: Prometheus & Grafana Setup

The concrete Prometheus + Grafana setup for LLM agents: token-flow metrics, cost counters, quality-drift and context-inflation alerts, and a four-row dashboard layout — with config snippets.

Ashot GrigoryanOctober 30, 2026

LLM Cost Optimization for Agent and RAG Workloads — article cover

6 min read

AI AgentsLLMCost

LLM Cost Optimization for Agent and RAG Workloads

A measurement-first approach to LLM cost: per-span cost attribution, input-token levers (retrieval filtering, prompt caching, history compaction), cascade routing validated by evals, and cost as a regression metric.

Ashot GrigoryanOctober 29, 2026

Load Testing LLM Applications: Why RPS Doesn’t Work — article cover

7 min read

PerformanceLoad TestingLLM

Load Testing LLM Applications: Why RPS Doesn’t Work

LLM load testing needs a different model: token flow instead of RPS, rate-limit cliffs instead of queues, and quality sampling per stage. A practical methodology with metrics, stages, and failure modes.

Ashot GrigoryanOctober 27, 2026

k6 vs JMeter vs Gatling vs Locust: 2026 Guide — article cover

7 min read

PerformanceLoad TestingTooling

k6 vs JMeter vs Gatling vs Locust: 2026 Guide

Honest comparison of k6, JMeter, Gatling and Locust by CI/CD fit, throughput per node, scripting language and load model — plus why the load profile matters more than the tool.

Ashot GrigoryanOctober 27, 2026

Internal Evals: Turning Production Failures Into a Regression Suite — article cover

3 min read

AI AgentsEvalsLLM

Internal Evals: Turning Production Failures Into a Regression Suite

How to build agent evaluation infrastructure that predicts production — harvesting tasks from escalations, mechanical grading, honest statistics, and the bridge to simulation.

Ashot GrigoryanOctober 19, 2026

Event-Driven Agents: Waking Safely in a World That Doesn’t Wait — article cover

3 min read

AI AgentsArchitectureLLM

Event-Driven Agents: Waking Safely in a World That Doesn’t Wait

From request-response to resident systems — event normalization, safety points, cooperative cancellation, preemption priorities, and the fast/slow path split for AI agents.

Ashot GrigoryanOctober 19, 2026

Fine-Tune, RL, or Just Fix the Prompt: A Decision Tree for Agent Teams — article cover

3 min read

AI AgentsFine-tuningLLM

Fine-Tune, RL, or Just Fix the Prompt: A Decision Tree for Agent Teams

A diagnostic sequence before any training run — SFT clones, RL generalizes, and why data and reward design beat algorithms. With the cheap-fix ladder in between.

Ashot GrigoryanOctober 19, 2026

Context Compression for Long-Running Agents: The Five-Layer Cake — article cover

3 min read

AI AgentsContext EngineeringLLM

Context Compression for Long-Running Agents: The Five-Layer Cake

Why lossy summarization breaks agent sessions — task-aware compression, the production layer cake, KV-cache conflicts, and why isolation beats compression.

Ashot GrigoryanOctober 19, 2026

Agent Memory: Four Tiers from a Text File to Fine-Tuned Weights — article cover

3 min read

AI AgentsMemoryLLM

Agent Memory: Four Tiers from a Text File to Fine-Tuned Weights

A practical memory architecture for AI agents — working notes, structured profiles, episodic retrieval, and learned parameters, plus the hygiene rules that keep memory from rotting.

Ashot GrigoryanOctober 12, 2026

RAG That Survives Production: Hybrid Retrieval, Reranking, and Recall Budgets — article cover

4 min read

AI AgentsRAGLLM

RAG That Survives Production: Hybrid Retrieval, Reranking, and Recall Budgets

Beyond naive vector search: chunking decisions, BM25 + dense fusion with RRF, rerankers, summary cards for statistical questions, and recall@k as the acceptance metric.

Ashot GrigoryanOctober 5, 2026

Prompt Injection in Tool-Using Agents: Defense in Layers That Actually Holds — article cover

4 min read

AI AgentsSecurityLLM

Prompt Injection in Tool-Using Agents: Defense in Layers That Actually Holds

Direct, indirect, and memory injection patterns in AI agents — source labeling, role hygiene, least-privilege tools, proposer-reviewer approval, and a hardening checklist.

Ashot GrigoryanOctober 5, 2026

Agent Tool Design: Descriptions as Prompts, and the Hundred-Tool Problem — article cover

4 min read

AI AgentsMCPLLM

Agent Tool Design: Descriptions as Prompts, and the Hundred-Tool Problem

How to design AI agent toolsets — perception vs execution vs collaboration tools, description contracts, MCP plumbing, progressive disclosure and tool search past 100 tools.

Ashot GrigoryanOctober 5, 2026

The Coding Agent Is the Blueprint: Files, Verifiers, and Failure Recovery — article cover

3 min read

AI AgentsCoding AgentsLLM

The Coding Agent Is the Blueprint: Files, Verifiers, and Failure Recovery

Why coding agents are the reference architecture for all AI agents — file systems as memory, tests as reward functions, four-layer failure taxonomy, and harness-first design.

Ashot GrigoryanOctober 5, 2026

Agent Observability: Traces, Cost Attribution, and Catching Loops in Real Time — article cover

4 min read

AI AgentsObservabilityLLM

Agent Observability: Traces, Cost Attribution, and Catching Loops in Real Time

How to make AI agents debuggable — execution trace trees, the five metrics that matter, loop detection alerts, and feeding production failures back into evals.

Ashot GrigoryanOctober 5, 2026

Context Engineering: The Discipline That Makes AI Agents Production-Ready — article cover

4 min read

AI AgentsContext EngineeringLLM

Context Engineering: The Discipline That Makes AI Agents Production-Ready

Why production agents fail over context, not intelligence — cache-aware prompt structure, layered information lifecycles, and isolation over compression.

Ashot GrigoryanSeptember 30, 2026

Agent Evals: Pass@k Is a Demo, Pass^k Is a Product — article cover

6 min read

AI AgentsEvalsLLM

Agent Evals: Pass@k Is a Demo, Pass^k Is a Product

Pass@k vs Pass^k explained: formulas, a reference table from p=50% to 95%, and how to choose the right agent evaluation metric from retry economics. A 60% step gives 99% Pass@5 but 8% Pass^5.

Ashot GrigoryanSeptember 30, 2026

When Multi-Agent Architecture Pays Off (and When It Just Burns Tokens) — article cover

3 min read

AI AgentsArchitectureLLM

When Multi-Agent Architecture Pays Off (and When It Just Burns Tokens)

A practical decision framework for multi-agent AI systems — information isolation over team metaphors, orchestrator topologies, and the metrics that keep costs honest.

Ashot GrigoryanSeptember 30, 2026

Building Production-Ready AI Agents: Lessons from Real-World Deployments — article cover

4 min read

AI AgentsLangGraphLLM

Building Production-Ready AI Agents: Lessons from Real-World Deployments

Practical insights from deploying AI agents in production — architecture patterns, tool integration, error handling, and the LangGraph framework.

Ashot GrigoryanApril 25, 2026

Performance Engineering at Scale: From Detection to Resolution — article cover

3 min read

Performance EngineeringMonitoringBackend

Performance Engineering at Scale: From Detection to Resolution

How we approach performance engineering in production systems — from identifying bottlenecks to implementing scalable solutions that handle millions of requests.

Aleksei SvetlovApril 20, 2026

Stay Updated

Subscribe to our RSS feed to get the latest articles on performance engineering, AI agents, and modern software development.

RSS Feed

Want to discuss your project?

Let's explore how our expertise in performance engineering and AI agents can deliver measurable results for your business.