Nitish Agarwal

Multi-agent systems are distributed systems, and most of the old lessons carry over: timeouts, retries, partial failure, ordering. The new twist is a component that can answer the same request two different ways. Which microservices habit has helped your agents most?

Stop training forecast models. Time series just went zero-shot and it runs on a CPU. #timeseries #ai #timeseries-ai #misteral-ai

Our audit agent went 3 weeks without escalating once. Then a spot-check found 11 bad approvals. "Ask if unsure" isn't an escalation path. Gate actions by risk tier, escalate async, staff the queue, and turn every human answer into an eval case. #AI #AgenticAI #AIDevelopment #SoftwareDevelopment

🚗⚡ Introducing voltbase — open-core tooling for EV charging data. Normalize OCM, OSM, OCPI & DATEX II → one model → API + MCP. Open data only. Provenance + licensing built in. Stale data is explicit. Apache-2.0. github.com/nitishagar/v... #EV #OpenData #MCP #OpenSource

Retries are load multipliers. The SRE book's 64-attempt cascade, rediscovered by every team that wired an LLM into a while-loop. Each retry re-sends a longer transcript: cost climbs while your odds fall. #AIEngineering

#AI #AIModels Your agent works in the demo. Works in staging. Fails some % in prod and the team argues about which model to swap to. That error rate is four failure modes wearing one lab coat. The 3-tier harness that tells them apart:

In 2010, Ellison said his vision was IBM's 1960 vision. In 2026 it's everyone's: Nvidia sells factories, AMD ships $5M racks, Huawei brute-forces supernodes. #AIEngineering

Feed an agent the same request twice and you may get two different answers one right, one confidently wrong. Your monitoring dashboard can’t tell the difference. A field guide to agent SLOs: nitishagar.medium.com/your-agent-h...

EU AI Act enforcement begins Aug. 2. If an auditor asked, "Can you prove what your model said last month?", could you? receipta: open-source, Ed25519-signed receipts for every LLM call. (github.com/nitishagar/receipta) #AI #AIGovernance #OpenSource

Agent memory is 2026's vector database: $24M raises, benchmarks the vendors dispute with each other, and a markdown file quietly doing the job. Claude Code and Codex already ship memory as files. Full argument, check the story below.

Everyone's focused on AI agents communicating. Far fewer are asking how agents prove who they are. ANS takes a pragmatic approach: build on DNS, PKI, and transparency logs instead of inventing a new trust system. A direction for secure agent-to-agent interactions. #AIAgents #SoftwareArchitecture

1.4M notebooks. 4% reproduced. ~116 papers human-verified in six years. Reproducibility is a labor problem so I built agents that rerun your science and show their receipts: github.com/nitishagar/r...

"Keep Opus out of the loop; it's 19x Haiku." That gap is 5x now. Tiered agents still cut bills 40-60%, but not because of price: token volume drives ~80% of agent performance, model choice ~5%. Why ↓ #AIAgents #LLMOps #AI

Most production AI agents are amnesiacs waking up, executing, and forgetting everything by the next session. But a massive architectural shift is happening. In my latest article, I break down the 8 major architectural decisions you now own when building agents. #AI #AIAgentArchitecture #AIMemory

Most multi-vertical AI agents fail because specialization is treated as a prompt problem. What worked for us instead: * template-scoped personas * composable prompt fragments * scaffold-aware reasoning One agent implementation. Multiple specialized behaviors. No prompt spaghetti. #AI #LLM #Agents

The "Agent Stack" just grew up. 🤖 Six months ago, we were worried about transport and caching. In 2026, it's about context economics and the three-layer stack: Skills, MCP, and Code Execution. #MCP #AI #LLMs #Engineering

Traditional SEO is dead for dev tools. The new game is "AI Retrieval Architecture." If you aren't shipping an llms.txt index and raw markdown twins for agents to fetch, you are practically invisible to modern AI. Stop optimizing for the crawl and start optimizing for the fetch. #SEO #LLM #WebDev

Open WebUI is a platform, not just a UI. If you're putting real streaming LLM agents behind it, you have to look past the demo and design for scale. I wrote down the architectural decisions we need to make before your first production outage. 🛠️👇 #LLM #OpenWebUI #AIArchitecture #Engineering

Building AI apps? You've probably hit the LLM streaming wall 🧱 BAML solves this beautifully by parsing broken LLM streams into typed, partial objects in real-time. Stop waiting for the final closing brace to render your UI! 🛠️✨ #SoftwareDevelopment #AI #LLMs #WebDev #Engineering #Tech

Finding vulnerabilities in CI is too late. You're already context-switching. 🛑 I wrote about a zero-code (just Markdown!) multi-agent architecture that moves security review to pre-commit. #DevSecOps #AIagents #AppSec Adversarial AI "validators" kill hallucinations before devs ever see them. 👇

AI coding tools optimize for green tests, even if the logic is broken. It’s time to bring back TDD as an AI constraint. 🦋 #SoftwareEngineering #AI #TDD #CleanArchitecture #LLMs #Programming #DevOps #SoftwareTesting

AI crawlers hit 50 billion requests/day. Most websites serve them an empty div. I built SoberAI an open-source Lighthouse for AI agent readiness. Give it a URL, get a score across SSR, schema, semantics, & content extractability. Free. No API keys. 🔗 github.com/nitishagar/sober-ai

We’re obsessed with the "Microservices" model. But in production, this leads to massive latency and "context drift." The fix? The Actor approach. Why? ✅ Deterministic routing ✅ Session-first state ✅ Actor semantics #AI #LangChain #CrewAI #OpenClaw #BuildInPublic

I wrote about "Agentic Debt" and the new OWASP LLM07:2025 classification. The industry is finally converging on a standard: Prompts are not secrets or config. #SoftwareEngineering #AI #LLM #DevSecOps #AgenticDebt #BuildInPublic

The 3 patterns for integration I cover in the article: 1️⃣ Webhook-based ingestion (The MVP) 2️⃣ Streaming detection via Kafka (Real-time) 3️⃣ RAG-based context retrieval (The Game Changer) #SRE #AIOps #DevOps #Engineering medium.com/p/ai-agents-...

Your AI agents will be compromised. The question isn’t “if,” but “what happens next.” In Part 2 of my Zero-Trust for AI Agents series, I cover: 🔹 Micro-segmentation to limit lateral movement 🔹 Runtime containment to stop blast radius spread #AIsecurity #ZeroTrust #CloudNative #AIOps t.co/sBTsmTEoHE

Your security stack wasn't built for AI agents. It's time for a Zero-Trust approach. I'm starting a series on securing AI agents. Part 1 is all about the foundation: IDENTITY. #AI #Cybersecurity #ZeroTrust #DevSecOps #AIAgents Link: medium.com/p/zero-trust...

Meta's LLaMA training failed every 3 hours across 16,384 GPUs. GitHub Copilot missed critical security bugs. Yet they recovered. In my deep dive I look at How OpenAI, Anthropic & Notion build evals that actually work with failures, fixes & frameworks you can use today. #AIEvals #MLOps #AI

#pyconau @nitishagar.bsky.social Building Production-Ready MCP Servers: From Protocol to Patterns http://youtu.be/Lrk-ZlaBYfY

REST APIs are dying for AI applications. It's not just about speed, it's about enabling AI agents that can maintain state, recover from failures, and coordinate complex multi-tool workflows in real-time. #LLM #Streaming #AIEngineering #MCPNews Full Story: nitishagar.medium.com/the-streamin...

Just wrapped up my talk at PyCon AU 25 on "Building Production-Ready MCP Servers" 🎯 We're moving from "every AI needs custom integration" to "build once, connect everywhere." Huge thanks to the @pyconau.bsky.social community for the engaging discussions! #PyConAU25 #AI #MCP #Python #TechTalks

The problem: Requirements get lost in translation Business need. In my article below I look at AI first requirements as code tools to solve this. #SoftwareEngineering #AIinDevelopment #TechnicalLeadership #DevOps #AgileTransformation #DeveloperProductivity Full guide: medium.com/p/the-markdo...

🐍✨ Excited to present on Model Context Protocol (MCP) at #PyConAU2025! Join me as we explore MCP patterns for better AI tool integration and making LLMs more powerful. 🎟️ Get your tickets here: pretix.eu/pyconau/2025... See you in 🇦🇺! #Python #AI #MCP #MachineLearning #Conference

Tired of state management hell in your AI agents? The solution has been hiding in your toolchain all along. A blueprint for building truly stateful LLMs using the power of continuations. #AI #SystemDesign #LLM #DeveloperExperience nitishagar.medium.com/the-architec...

In my recent article I explore how "SF BOLAA" pattern can be blueprint for AI microservices going forward. #AI #Microservices Link below: medium.com/@nitishagar/...

The magic of your LLM demo wears off fast in production. 😟 Hallucinations, regressions, silent failures... The answer isn't just better prompts—it's a better process. It's time for LLM Evals in your CI/CD pipeline. nitishagar.medium.com/from-magic-t... #LLM #AI #MLOps #CICD #DevOps #GenAI

Stop writing every line of code, start conducting it. My latest article breaks down the practical agentic workflow that turns AI into your junior dev, making you a true 10x engineer. #AICoding #SoftwareEngineering #DeveloperProductivity #FutureOfWork

Building in AI? Decide if you're shipping the "picks & shovels" (MCP server platform) or the "gold" (AI Agent) – your strategy depends on it! #AI #ProductStrategy #MCP #AIAgents #TechInnovation

Wrote an article on lightweight models that can run on standard hardware, significantly reducing complexity from a DevOps and infrastructure standpoint. #infra #ai #tech Why Smaller, Efficient Models Are a Game Changer for DevOps & Infra medium.com/@nitishagar/...

DeepSeek: A Disruptor in the AI Cost Equation The article examines how DeepSeek serves as a catalyst for companies to explore cost disruption through AI tools and augmentation. #ai #ai-strategy #aws #software-strategy

Just published article on potential use of serverless with LLM: medium.com/@nitishagar/... #llm #serverless #ai

Jailbreaks are a major AI risk. Anthropic's innovative defense is a wake-up call for enterprise security. Get the breakdown + actionable steps: link.medium.com/pLk4lHCsJQb #AISecurity #LLMSecurity #JailbreakAI #Anthropic #EnterpriseAI