标签：#Agents

Originally published at kunalganglani.com — read it there for inline code, hero image, and live links. Generative AI vs agentic AI vs AI agents. Three terms, used interchangeably by people who should know better, burning engineering budgets across the industry in 2026. Generative AI refers to models that produce new content — text, images, code — from a prompt. AI agents are software systems that wrap those models with planning, memory, and tool use to pursue goals autonomously. Agentic AI is the broader paradigm: orchestrated systems of agents, workflows, and decision-making that operate with minimal human oversight. Getting these distinctions wrong doesn't just lose you a Twitter argument. It determines whether your production system costs $500/month or $50,000. Every quarter, someone on a leadership team says "we need to go agentic." What they usually mean is one of three completely different things. And the architecture you pick for each one has wildly different implications for cost, latency, reliability, and maintenance burden. I've watched teams burn entire quarters building autonomous agent systems when a well-tuned prompt engineering pipeline would have shipped in a week. That's not a hypothetical. I watched it happen twice in 2025. This post cuts through the buzzword soup. I'll define all three paradigms with concrete technical distinctions, show you how they map to real production architectures, and give you a decision framework for picking the right one. What Is Generative AI? The Engine, Not the Vehicle Generative AI is the foundation layer. It's a large language model (or image model, or audio model) that takes an input and produces new output. GPT-4, Claude, Gemini, Llama — these are all generative AI. You send a prompt, you get a completion. That's it. The critical thing to understand: generative AI is stateless by default . Each API call is independent. The model doesn't remember what you asked five minutes ago. It doesn't plan a sequence of steps.

If you've ever felt like LangChain was too heavy, you're not alone. The dependency tree is enormous. Abstraction layers pile up. At some point you lose track of what's actually happening underneath. That frustration has pushed a lot of people toward lighter alternatives — frameworks that prove you can build a capable agent without a hundred transitive dependencies. Agno is one of those alternatives. It started as Phidata and rebranded in early 2025. I spent an afternoon installing Agno v2.6.17 in a clean sandbox and running through Calculator tools, Wikipedia retrieval, Pydantic structured output, and a two-agent Team. I'll share the real execution logs and, more importantly, the traps I hit that the docs don't warn you about. What Agno Is and Where It Came from Phidata built a solid reputation as "the Python framework for AI assistants." When it rebranded to Agno in 2025, the design philosophy got articulated more clearly around three ideas. Model-agnostic from day one. Over 70 LLMs — OpenAI, Anthropic, Google, Ollama, Cohere — can plug in with the same code structure. Swap the model, keep the agent logic. Multimodal as a default. Text, image, audio, video agents all use the same API surface. You don't need a different abstraction layer for each modality. Multi-agent orchestration as a first-class citizen. The Team class is built in. You can switch between coordinate , route , and collaborate modes with a single parameter change. Reading that, I thought: "How is this different from LangChain?" The answer showed up when I actually wrote code. Agno favors composition over class inheritance. One agent takes about 6 lines to set up. There's far less boilerplate to wade through. Installation: No Dependency Hell pip install agno google-genai ddgs wikipedia The agno package installs just the core. Tools require their own extra dependencies — wikipedia for the Wikipedia tool, google-genai for Gemini. This lazy-loading approach keeps the base install clean. $ python3 -c "im

Google Open Knowledge Format: Why Enterprise Agents Need a Knowledge Layer, Not Just More Tools Most enterprise AI conversations still start in the wrong place. They start with the model. Which model should we use? Which framework should we adopt? Which vendor has the best agent platform? Which tools should we connect next? These are fair questions. But in real enterprise architecture, they are not the hardest questions. The harder question is this: Can our AI systems actually understand how our business works? That is why Google Cloud’s article on Open Knowledge Format caught my attention. The article talks about a simple but important idea: representing knowledge in a way that humans can read and machines can use. In OKF, that means markdown for the content and structured metadata for context. At first glance, that may sound too simple. But that simplicity is the point. Enterprises do not need another place where knowledge goes to die. We already have enough portals, catalogs, wikis, dashboards, folders, and internal tools. What we need is a practical way to package knowledge so it can be reviewed, versioned, governed, searched, and reused by both people and AI agents. That is where this idea becomes very relevant for agentic AI. The Real Enterprise AI Problem Most organizations already have the knowledge their AI agents need. They have it in databases, dashboards, tickets, architecture notes, runbooks, Confluence pages, data catalogs, code comments, incident reports, old project documents, and the heads of experienced employees. The issue is not that knowledge does not exist. The issue is that it is fragmented. Some of it is outdated. Some of it is duplicated. Some of it is tribal. Some of it is locked inside tools. Some of it is written for humans but not structured enough for AI systems to use reliably. This becomes a serious problem when we move from AI assistants to AI agents. An assistant can give a helpful answer. An agent does more. It plans, selects tools

Microsoft recently introduced at Build 2026 Microsoft Scout, an always-on agent. Scout belongs to a new category of agents Microsoft called Autopilots: always-on agents that work autonomously on a user’s behalf with their own identity, without needing to be prompted each time. Microsoft Scout integrates with Work IQ and is based on the open-source agent framework OpenClaw. By Bruno Couriol

你知道吗？GitHub 上有一个 17,889 Stars 的开源项目，能让你的 AI Agent 拥有跨会话的持久记忆——不是简单的向量检索，而是一个会自动进化的知识图谱。但大多数开发者只用它来做基础的文档搜索，完全忽略了它真正的能力。 Cognee 是一个开源 AI 记忆平台，它把知识图谱、向量搜索和认知本体论生成统一到一个记忆层中。在 2026 年，AI Agent 正从单轮对话机器人向长时间运行的自主系统演进，而瓶颈不再是模型能力，而是上下文管理。以下是大多数人不知道的五个隐藏用法。 隐藏用法 #1：自动图谱同步的会话记忆 大多数人的做法：把对话历史存在简单的列表或向量数据库里，上下文长了就塞进 prompt。这在前几轮还行，但会话一长就迅速退化。 隐藏技巧：Cognee 的会话记忆充当快速缓存，会在后台自动同步到持久化知识图谱。你既能获得内存上下文的速度，又能拥有图数据库的持久性——而且同步过程完全不需要手动编排。 import cognee import asyncio async def agent_session (): # 会话记忆——快速、临时、按会话隔离 await cognee . remember ( " 用户询问了 Q3 收入趋势并请求导出 CSV。 " , session_id = " support_ticket_4421 " ) # 后续查询会话记忆（快速路径） results = await cognee . recall ( " 用户问了什么收入相关的问题？ " , session_id = " support_ticket_4421 " ) # 会话结束时，会话记忆自动同步到永久图谱 # 无需手动导出，不会丢失任何数据 asyncio . run ( agent_session ()) 效果：你的 Agent 在会话内部保持对话上下文以实现快速响应，但会话结束后不会丢失任何信息。知识图谱会自动跨所有会话积累洞察。 数据来源：Cognee GitHub 17,889 Stars，README 文档中 session_id 参数和自动同步行为在"Use with AI Agents"章节有详细说明。 隐藏用法 #2：面向领域推理的本体论 grounding 大多数人的做法：把文档灌入向量数据库，依赖语义相似性做检索。模糊匹配还行，但当你需要结构化的、领域感知的推理时就不行了。 隐藏技巧：Cognee 的 cognify 流水线不只是嵌入文档——它会从你的数据中生成认知本体论，创建带有类型化关系的结构化知识图谱。这意味着你的 Agent 可以对实体及其连接进行推理，而不仅仅是找到相似的文本。 import cognee import asyncio async def build_domain_memory (): # 摄入领域文档 await cognee . remember ( """ 客户 Acme Corp 有 3 个活跃订阅。 订阅 A：企业计划，到期日 2026-09-15。 订阅 B：入门计划，已于 2026-03-01 到期。 客户经理是 Sarah Chen。 升级路径：Sarah -> 销售副总裁 -> CRO。 """ ) # Cognee 自动提取实体和关系： # (Acme Corp) --拥有--> (订阅 A) # (订阅 A) --类型--> (企业计划) # (订阅 A) --到期--> (2026-09-15) # (Sarah Chen) --管理--> (Acme Corp) # 现在进行结构化精确查询 results = await cognee . recall ( " 哪些客户在未来 90 天内订阅到期？ " ) # 返回 Acme Corp 及具体订阅和日期—— # 而不仅仅是"关于订阅的相似文本" asyncio . run ( build_domain_memory ()) 效果：你的 Agent 不再依赖向量相似性碰运气，而是从理解实体类型、关系和时间约束的本体论中获得结构化答案。 数据来源：Cognee README "Product Features"章节描述了"ontology grounding"和"cognitive-science-grounded ontology generation"；ArXiv 论文 2505.24478 关于优化知识图谱与 LLM 的接口。 隐藏用法 #3：通过共享图谱实现跨 Agent 知识共享 大多数人的做法：每个 Agent 维护自己独立的记忆。客服 Agent 无法受益于销售 Agent 昨天学到的东西。知识按设计被隔离。 隐藏技巧：Cognee 的知识图谱是一个共享基础设施层。多个基于不同框

Here is the detailed, deep-dive article tailored for DEV.to, written in a natural, highly technical style, completely free of icons, and designed to resonate with developers building agentic workflows. Building an Autonomous AI Experience Engine: Taming the Multi-Agent CLI Fleet As developers integrate more AI tools into their workflows, a new architectural problem has emerged: agent sprawl. We have incredible tools like Claude, Grok, and Codex running in our terminals, but they operate in silos. They lack shared memory, they step on each other's toes, and coordinating them feels like herding cats. To solve this, I built TechSphereX Studio — an open-source, polyglot AI Experience Engine. It is an autonomous multi-agent platform designed to intercept AI coding actions, orchestrate goal-driven work across a fleet of CLI agents, and mathematically learn from every session to improve future outcomes. Here is a deep dive into how I moved from isolated prompt engineering to a fully automated, self-learning agentic brain. Key Architectural Pillars 1. The 3-Layer Intercept Pipeline Before any CLI executes a command, TechSphereX intercepts the action to determine if the system already knows how to solve the problem based on past experiences. This happens across three highly optimized layers: Layer 1 (Read-only Filter): Evaluates the action in under 1ms. If the action is non-destructive (like a simple read), it skips heavy processing to save resources. Layer 2 (Semantic Search): Uses Qdrant running locally to perform vector embeddings and search the system's history for similar past tasks in under 50ms. Layer 3 (LLM Rerank): Passes the semantic results to a local Ollama instance to filter out contextually irrelevant data in under 500ms, ensuring the execution agent only receives high-fidelity context. 2. The Agentic Brain & Multi-Role Teams Instead of throwing a massive, complex prompt at a single coding agent, TechSphereX mimics a multi-role engineering team. The pipeline st

Tokenmaxxing was the hottest trend in Silicon Valley earlier this year, with CEOs encouraging employees to push AI usage as far as it would go. Then the bill came due. Uber reportedly blew through its annual AI budget in a few months, some companies cut Claude licenses for parts of their org, and Meta killed its internal leaderboard. This tension between […]

Nvidia's self-improvement program for robots enlists teams of AI coding agents.

Last month my OpenClaw agent kept making the same mistake: it would run a health check, the script would fail silently, and the agent would report "all systems operational" with total confidence. It wasn't broken. It was just doing what it was built to do — execute tasks — without any mechanism to learn from the outcome. So I built it a self-improvement loop. Every night at 2 AM, an isolated OpenClaw session wakes up, reads the previous day's execution logs, identifies patterns in what went wrong, and updates the agent's memory files. No human in the loop. No re-deployment. Just... learning. Here's what I built, what broke, and what actually works. Why Self-Improvement Is Hard for Personal Agents Enterprise AI labs solve this with massive infrastructure: reinforcement learning pipelines, full fine-tuning jobs, A/B testing frameworks that run for weeks. For a personal agent running on a cron job, that's not an option. The self-improvement loop for a personal OpenClaw setup has to be lightweight. It has to run in seconds, not hours. It has to write to plain text files that the next session will actually read. And critically, it has to avoid the feedback loop problem — an agent that rewrites its own improvement logic can spiral into nonsense if there's no anchor. The key architectural decision I made: separate the executor from the critic . Your main agent runs tasks. A separate isolated session reviews what happened and recommends changes. The main agent applies them on the next run. No single session is both judge and executioner. The Nightly Cron: What Actually Runs This is the cron I have running at 2 AM ET every morning: { "name" : "nightly-self-improvement" , "schedule" : { "kind" : "cron" , "expr" : "0 2 * * *" , "tz" : "America/New_York" }, "sessionTarget" : "isolated" , "payload" : { "kind" : "agentTurn" , "message" : "Review the last 24 hours of OpenClaw execution. Read memory/$(date +%Y-%m-%d).md and memory/yesterday.md. Identify 3 patterns where the agent u

Uber recently described an internal architecture for propagating identity across multi-agent AI workflows. The design aims to perserve user context, agent provenance, and scoped access as agents delegate work and call internal tools. The case study aligns with Auth0’s view that AI agents need permissions based on delegated authority, scoped credentials, and explicit human approval boundaries. By Eran Stiller

Originally published on the Sieve blog . Sieve is an open-source (Apache 2.0) context-reduction proxy — I work on it, and I've tried to keep this post about the problem rather than the tool. If you run an LLM agent for real work, the bill is the part nobody warned you about. It starts small, it grows with use, and the worst of it is invisible — most of what you pay for on any given turn is text the model has already seen, or text you never meant to send. There's a temptation to treat this as one problem with one fix. It isn't. An agent's token bill has two distinct causes, and they need two genuinely different kinds of tool. This post is about telling them apart — because once you can, the question stops being "which tool wins" and becomes "which of my two problems am I looking at right now." The bill is mostly things you didn't choose Start with where the tokens actually go, because it's rarely where people assume. When your agent calls a tool, the model doesn't just pay for your request — it pays for the machinery of asking. Anthropic's own pricing documentation spells this out: the tools parameter alone adds hundreds of tokens of schema to every request, the bash tool adds a fixed overhead, and a single web fetch pulls the fetched page straight into your context — "Average web page (10 kB): ~2,500 tokens... Research paper PDF (500 kB): ~125,000 tokens". A tool result you glance at once and never need again can cost more than the entire conversation around it. Now add the part that repeats. On every turn, a typical agent re-sends its system prompt, its full tool catalogue, its persona, and the conversation so far. The variable part of the request — what you actually typed — is often the smallest thing in the payload. The fixed overhead, multiplied across every turn of a long session, is the bill. So the cost has two shapes, and they're not the same shape: Verbose machine output — JSON tool results, logs, search dumps, fetched pages, code listings. Big, one-off, an

GitHub has introduced the GitHub Copilot app, a desktop control centre for agent-native development that aims to keep engineers in charge while AI agents handle more coding work. Mario Rodriguez writes on the GitHub blog that the recent wave of coding agents has brought faster delivery but also "disjointed workflows, more context switching, and too much time spent reviewing agent-generated code". By Matt Saunders

This post is a TypeScript implementation of the pattern described in "Beyond the Agentic Loop: The Orchestrator Pattern for Multi-Agent Systems" by Amogh Ubale (Stackademic). The original is Python with generic agents; here we keep the idea intact and re-theme it as a shopping assistant so the three execution modes have something concrete to chew on. All the design credit goes to that article — go read it first. The cast: a handful of shopping agents Before the pattern, the scene. The demo is a small storefront assistant backed by a few single-purpose agents: Catalog — list the categories on offer, or search products by keyword and price. Inventory — check stock and availability for a product. Pricing — look up the current price and any active promotions. Reviews — fetch a product's rating and review highlights. Order — place an order for a product. A customer request might need just one of these, several of them at once, or a few in a strict order — and deciding which of those shapes a request calls for is exactly what the orchestrator is for. The problem: the LLM as a while loop The default way to build a multi-agent system is the agentic loop : you hand the model a bag of tools and let it drive. think → call a tool → observe the result → think again → call another tool → … The LLM is both the brain and the control flow. That's wonderfully flexible, and it's the right tool when the task is open-ended and you genuinely don't know the steps in advance. But in production it has three nasty properties: Unpredictable shape. Every "think" step is another LLM round-trip, so a three-agent task might take 3 calls or 9 — you don't know until it runs, and latency swings with it. (The article clocks a representative three-agent query at ~7 calls through the loop; the wall-clock and spend follow, but the unpredictability is the part that actually bites.) Non-determinism. The same question can take a different path each time, which makes behavior hard to reason about and hard t

You've built the agent, the reply loop works, the demo lands — and then someone asks the question you didn't budget time for: "so what address does it send from?" Suddenly you're staring at noreply-svc-prod2@yourcompany.com realizing that the first thing every recipient sees isn't your prompt engineering. It's the From line. An agent's email address is an interface. Humans parse it before the subject, mail servers judge it before the body, and spam filters score it before any human sees it at all. Once your agent has a real mailbox — which is what Nylas Agent Accounts (currently in beta) provide — address design becomes a product decision with three layers: the local part, the domain, and the disclosure question. The local part: role beats persona beats hash Take three candidate addresses for the same scheduling agent: scheduling@ — a role. Tells recipients what the mailbox does and implies that emailing it is how you use the service. jane.ai@ — a persona. Friendlier in a sidebar avatar, but it invites recipients to treat the sender as a colleague, with all the expectations that carries. bot-7f3a@ — an artifact. Screams "auto-generated," gets mentally filed next to spam, and gives a human nothing to anchor on. The docs consistently model the first pattern — sales-agent@ , support@ , scheduling@ appear throughout the Agent Accounts overview — and I think that's right for a reason deeper than convention. A role address makes an honest promise about capability: scheduling@ claims it can schedule, nothing more. A persona address makes an implicit promise of general competence that current agents can't keep. When jane.ai@ fails to understand a simple request, it reads as a person being obtuse; when scheduling@ fails, it reads as a tool hitting its limits. Same failure, different trust damage. The overview's framing is worth internalizing: an agent identity should be like any other user in your organization — reachable, persistent, accountable. Address design is how that

Day 1, 4:02 p.m.: a customer asks your agent a billing question and gets an answer. Day 6, 9:30 a.m.: they reply "actually, that didn't work." If your agent lives in a chat widget, that second message starts from zero — the session died with the tab, the context is gone, and the customer gets to repeat themselves. If your agent lives in a mailbox, the reply arrives inside the conversation , with the full history attached by the protocol itself. That's the argument in one before/after: chat sessions evaporate; email threads persist. And for agents that work across days rather than minutes, the thread is the most underrated memory substrate available. The protocol already built your memory layer Email threading runs on three headers, as the threading docs lay out. Every message carries a globally unique Message-ID . A reply adds In-Reply-To (the ID it's answering) and References (the full chain of IDs, oldest to newest). By the time a thread is five messages deep, References holds five Message-IDs in order — a complete, tamper-evident record of the conversation's shape, maintained by every mail client on earth. Compare that to what we hand-roll for chatbots: session stores, conversation tables, context windows we serialize and rehydrate. Email gives you the equivalent for free, federated across organizations, and — this is the part I find most compelling — human-auditable . Anyone with mailbox access can read exactly what the agent's memory contains, because the memory is the correspondence itself. No vector store inspection tools required. With Nylas Agent Accounts (in beta), the agent owns the mailbox where this accrues, and you never parse headers by hand. The Threads API groups messages by their header chain; each thread object gives you ordered message_ids , participants , and activity timestamps. When a reply fires message.created , the payload includes a thread_id — fetch the thread, walk its messages, and the agent has its full conversational past before decid

A human assistant borrows the boss's calendar; a scheduling bot owns its own. That one difference dissolves most of what makes calendar automation miserable. The borrowed-calendar model is how nearly every scheduling tool works today: connect to a person's Google or Microsoft account via OAuth, request calendar scopes, and act on their behalf. It works, but the seams show everywhere. The human's calendar fills with bookings the bot manages. Delegation permissions vary by provider and admin policy. Tokens expire when the person changes their password or leaves the company. And the bot has no address of its own — every invite, every confirmation email, appears to come from a person who didn't write it. Nylas Agent Accounts (currently in beta) invert this. Each account is a real mailbox and a real calendar , provisioned automatically, owned by your application. From a participant's perspective there's nothing special about it — it's just another attendee on the invite. What "owning availability" actually changes When the bot's calendar is its own, availability stops being a permissions question and becomes a query. The agent calls the free/busy endpoint against its own primary calendar, gets back busy blocks for a time window, and proposes open slots. No delegation, no scopes negotiation, no "the admin needs to approve calendar sharing for service accounts." The scheduling-agent tutorial wires the whole loop: a meeting request lands at scheduling@agents.yourcompany.com , a message.created webhook fires, an LLM parses duration and timezone, the agent checks its own free/busy and replies with 3 candidate slots. When the human picks one, the agent creates the event: curl --request POST \ --url "https://api.us.nylas.com/v3/grants/<GRANT_ID>/events?calendar_id=primary&notify_participants=true" \ --header "Authorization: Bearer <NYLAS_API_KEY>" \ --header "Content-Type: application/json" \ --data '{ "title": "Product demo", "when": { "start_time": 1744387200, "end_time": 174

Every project I touch lately ends up needing the same awkward thing: a reliable way to pull the web into a script or an agent. Not a brittle scrape held together with CSS selectors and hope, but something that takes a URL and hands back clean, structured text I can actually pipe into the next step. I have built that wrapper more than once, and it is never as small as you think it will be. So when Mozilla dropped the tabstack CLI, a single Go binary that wraps the Tabstack AI API, I wanted to spend a proper afternoon with it. The pitch on the README is direct: every web interaction your agent or stack needs, from the terminal or a script. It turns any URL into clean Markdown or schema-shaped JSON, runs natural-language browser automation, and answers research questions with cited sources. The part that made me sit up is that the output is pretty in a terminal and pipeable into jq without a flag. That is a small detail, and it tells you the people who built it actually live on the command line. Let me walk you through it the way I poked at it myself. Getting it installed There is no runtime to install and nothing to bootstrap, because it ships as a single static binary built for macOS, Linux, and Windows. The quickest route is the install script: curl -fsSL https://tabstack.ai/install.sh | sh That fetches the right binary for your platform and puts it on your PATH , and you are ready to go. If you would rather not pipe a script into your shell, there are a couple of alternatives. With Go on your machine you can use: go install github.com/Mozilla-Ocho/tabstack-cli/cmd/tabstack@latest That drops the binary in $GOPATH/bin , which is usually ~/go/bin . If your shell cannot find tabstack afterwards, you almost certainly have not got that directory on your PATH : export PATH = " $HOME /go/bin: $PATH " And if you want to avoid Go entirely, there are pre-built binaries on the Releases page, or you can clone the repo and run make install-local , which builds it and copies it t

The email mailbox just became an API resource, and that matters far more than the setup time it saves. For most of software history, a real email address — one that sends, receives, and threads — was an artifact of IT process. Someone created it in an admin console, someone else configured the client, and your application got access through OAuth consent screens and refresh tokens borrowed from a human. Compare that to how you get a database, a queue, or a TLS cert today: one API call, one ID back, done. Nylas Agent Accounts (currently in beta) close that gap. The quickstart goes from API key to a sending-and-receiving mailbox in under 5 minutes, and the provisioning step is a single request: curl --request POST \ --url "https://api.us.nylas.com/v3/connect/custom" \ --header "Authorization: Bearer <NYLAS_API_KEY>" \ --header "Content-Type: application/json" \ --data '{ "provider": "nylas", "settings": { "email": "test@your-application.nylas.email" } }' No refresh token, no OAuth dance — unlike OAuth providers, the "nylas" provider needs only an email address on a registered domain. The response contains a grant_id , and that one ID drives everything else: messages, drafts, threads, folders, attachments, calendar, webhooks. There are actually three ways to create an account — this API call, the Dashboard, or a single CLI command ( nylas agent account create ) — but they all end at the same place: a live mailbox. Why 5 minutes is a threshold, not a convenience Provisioning time isn't a linear cost. There's a threshold below which a resource changes category — from "thing you request" to "thing your code creates." Virtual machines crossed it with cloud APIs and we got autoscaling. TLS certs crossed it with ACME and we got HTTPS-by-default. Mailboxes crossing it means email addresses stop being scarce, pre-planned identities and start being something a program allocates when it needs one. What does that look like in practice? System mailboxes without ceremony. A support

I think a lot of people still imagine AI coding as opening ChatGPT, asking for code, and copy-pasting the result. That's not really how I work anymore. The biggest shift for me is that planning matters far more than coding. Earlier, execution was expensive, so most of the effort went into writing code. Now execution is cheap. I can have an agent implement something in minutes. The hard part is making sure the plan is correct. Most of my effort goes into thinking through the architecture, edge cases, failure modes, test strategy, and how the change fits into the broader system. If the plan is vague, the agent will confidently implement the wrong thing. The quality of the result is mostly determined by the quality of the plan. Once I have a plan, I break it into small independent pieces. Each piece should be executable without additional clarification. If an agent needs to stop and ask questions, the task probably isn't broken down enough. Those pieces become tickets. Then an agent picks up a ticket and implements it. The important thing is that the agent isn't operating in a vacuum. I try to give it a good environment to work in: Clear architectural rules Reusable skills and workflows Guardrails Hooks for things that must always happen One lesson that really stuck with me is that instructions are guidance, not guarantees. At one point I had "always use a git worktree" written in AGENTS.md. The model still ignored it occasionally. When I dug into it, the answer was simple: models can drift from instructions. So if something absolutely must happen, don't rely on instructions. Enforce it. Put it in a hook, script, validation step, CI check, or some other deterministic mechanism. If it is important, make it impossible to skip. Once the implementation is done, the agent opens a PR. This is where another useful pattern comes in: don't let the same model review the code it wrote. I usually have one model implement and another model review. Different models catch different t

Paul Klein discusses the distributed systems challenges of scaling cloud-hosted browser infra for AI agents. He explains how to manage bursty, stateful multi-tenancy and secure Chromium environments against remote code execution using Firecracker. He also shares how to leverage the Model Context Protocol (MCP) to turn complex websites into accessible agentic tools. By Paul Klein