Agent

What makes an Agent Skill valuable is not a clever prompt, but a clean division of responsibility: deterministic work goes to code, judgment goes to the model, and execution confirmation returns to the human. Structured contracts hold those parts together. I reached that conclusion by dissecting a storage-cleanup Skill that can delete local files from a web page. Deletion is one of the most consequential powers an agent-adjacent tool can expose. The design did not make me fearless; it gave me specific controls I could inspect before deciding whether to click. ...

After the Framework Stood Up, I Realized I Had No “Shared Workshop” The five-part Info-to-Creation series got the framework standing: information gets collected and denoised, records settle into half-finished goods, knowledge gets structured into capability, creation reorganizes it for an audience. But after writing that line in Layer 3: Knowledge — “your knowledge base is the workshop you share with your AI” — something kept nagging at me. ...

What actually happened here How large a team does it take to rebuild a four-year-old blog with 120+ posts — content architecture, operations, all of it? My answer, delivered over the first half of this year: one person and a crew of agents with clearly divided jobs. “Crew” is a working metaphor, not a fixed headcount of digital employees. As of July 2026, the repository contains seven Claude skills and four workflows whose names begin with seo-; Claude Code, Codex, the Claude jobs inside GitHub Actions, and temporary review sessions are started only when a task calls for them. The human owns judgment, boundaries, and the final signature. Agents own scoped work that can be tested. The result is the site you’re reading — cubxxw.com . It isn’t merely a redesign. It turned “a blog” from a static site into a system that generates a daily report, prepares fix proposals, and exposes its own search interface. ...

Suppose you had to place a bet today on one direction in AI agents. What is the first question you should ask? Most people ask whether the thing can be built. By mid-2026, that is often the least interesting part. The colder question is: why has nobody made this opportunity routine yet? The answer is a useful filter. If nobody has solved it because everyone just noticed it, competition is probably coming. If nobody has solved it because the work is dirty, slow, and carries consequence, there may be something defensible—or there may be no market. The rest of the article is how I try to separate the two. ...

Suppose you actually have one now — an agent that takes a job end to end. Pulls the data, writes the code, runs the tests, opens the PR, updates the docs. It doesn’t need you feeding it prompts line by line. You hand it the task at night and go to sleep. The real question isn’t whether it finishes. In coding, research, and content workflows, model capability is often already sufficient to produce a plausible result. That does not mean capability has stopped mattering everywhere: in unfamiliar domains and genuinely novel tasks, it can still be the limiting factor. But once an agent is capable enough to act, a different bottleneck appears — ...

The quiet dashboard can be the dangerous one A morning briefing with no new items may mean that nothing relevant happened. It may also mean an RSS route broke, an API began returning empty pages, or a credential expired during the night. From the reader’s side, those very different states look identical: silence. That is the first lesson of a personal intelligence system. Gathering more information is easy to demonstrate; knowing whether the machine is healthy, whether its summaries are faithful, and whether a signal deserves action is the real work. The system I want is not an account that publishes automatically. It is a pipeline that can show me what changed, why it may matter, what evidence supports the claim, and which decision—if any—should follow. ...

How many agents can one person afford to keep running? The wrong way to answer is to count agents. An “agent” might classify one paragraph, search for twenty minutes, or refactor a repository through eighty tool calls. The useful unit is not the agent. It is the successful task, with its input tokens, output tokens, tool charges, retries, and human cleanup attached. That distinction changed how I think about two related but different choices: low-cost hosted model APIs and open-weight models that can be self-hosted. Falling API prices matter, but “90% cheaper” is meaningful only when three things are visible: ...

A counterintuitive signal: it starts prompting you Start with a question. Suppose one day you open your workspace and the agent isn’t sitting there quietly waiting for your next command. It speaks first: “I noticed section three of yesterday’s proposal is still unfinished. I drafted something in the voice you used last week — want to take a look now?” Is that thoughtful, or is it presumptuous? Over the past six months I’ve noticed a counterintuitive signal while following agent launches, primary sources, papers, and benchmarks: more products are experimenting with agents that do not wait for the next instruction. They use available context, estimate what is worth doing, and bring a suggestion forward. ...

Suppose you ask an AI system to track one field — papers, releases, benchmarks, first-party changelogs, and the conversations around them. How far can it get on your behalf? This essay is not a census of every product. It is a field note from the systems I tested between January and June 2026, across AI research and developer-tool sources. In that sample, three patterns kept recurring: subscription aggregation, change monitoring, and agentic search. They were capable, but they shared a boundary: automation can haul information remarkably well; without your goals, context, and feedback, it cannot reliably make the final judgment for you. ...

A Year Later, Why Does AI Still Feel Like Extra Work? AI now writes copy, translates documents, summarizes meetings, and inspects spreadsheets. Yet one honest question cuts through the excitement: has it taken over a business step, or do you still explain the background, judge the answer, and decide what happens next every time? The model may be smarter. The person carrying the context and the responsibility often has not changed. ...

After Three Hours, What Is Left on the Table? I have had evenings that began with a modest intention: settle the angle of an essay. Soon I was discussing titles with AI, then business models, then the meaning of work. The conversation flowed beautifully. When I closed the window, the page was still blank. That does not make the conversation worthless. It reveals that exploration and execution are different kinds of work. ...

“It Runs” Is Not a Reliability Standard Most personal AI workflows begin with one acceptance test: did it produce something? A draft appears, a patch compiles, ten pages become one, and the task feels finished. After enough repetitions, however, the costly failures are rarely dramatic. They arrive quietly: a polished paragraph built on a stale source, a tool call that never completed, a plausible plan that solved the wrong problem. ...

A Second Brain Is a Working System, Not a Larger Notebook Most note systems are built for a future reader called you. You save a link, polish a heading, add two tags, and trust that one day you will return. Usually, you do not. An AI-native second brain begins with a different question: What useful work should this note make possible? The answer might be modest: turn an inbox note into a project brief, retrieve your earlier judgment before a meeting, or draft an article from claims you have already verified. The point is not to make AI read everything. The point is to give it a narrow, legible field in which it can help without quietly rearranging your life. ...

Growing Bigger, Getting Less Useful We’ve reached layer three. Information has been captured and denoised; records have been written down and polished into semi-finished products — now the question is: how do you turn these semi-finished products into actual knowledge? Let’s start with a definition. Knowledge is structured, repeatedly reusable material relevant to you: a mental model, a handful of skills, a methodology, along with your judgment, positioning, and values. Its keyword is reusability, and it solves your own problems. ...

open-lovable begins with a compact promise: give it a website, get back a React application you can keep editing. The generated page is the visible trick. The more useful subject is the plumbing underneath—how scraping, model output, and untrusted code execution are joined without surrendering the whole process to the model. First, pin the clock. This article audits firecrawl/open-lovable commit 69bd93bae7a9c97ef989eb70aabe6797fb3dac89 , dated November 19, 2025. “In the project” below means that exact revision. “The platform today” refers to current vendor documentation. A fixed commit is a small discipline, but it prevents later product features from being smuggled into an older implementation. ...

“Software is eating the world.” — Marc Andreessen, 2011 “Now AI is eating software—and the question for the rest of us is: what’s left for one human, alone, in front of a screen?” — me, asking myself one night in 2026. Prologue: How Big Does One Person Need to Be? In February 2026, I ran my first complete overnight agent. I set a prompt, dropped it into Claude Code in a loop, and went to sleep. At 7 a.m. the next morning, what I saw on the screen was: 6 commits, 4 PRs, 3 auto-rolled-back failures, and a research brief I hadn’t even read myself. ...

An architecture diagram is a promise. A code audit asks which parts of that promise have acquired weight. My first version of this article treated Relay as a public open-source proposal whose Agent layer had not yet been built. Both claims are now wrong. As of July 31, 2026, the public GitHub URL previously cited by this article returns 404, while I can inspect a private local checkout. This piece is therefore a private/local implementation audit, fixed to local commit 22586e17ccd43cfaff0512511e71a100c5341608. Readers should not assume that repository or commit is publicly downloadable. ...

Context engineering is the set of strategies for curating, ordering, and evicting the tokens that enter an LLM’s context window on each inference call—system instructions, retrieved documents, conversation history, tool definitions, and memory. The short version is this: prompt engineering optimizes the wording of an instruction; context engineering optimizes the wiring of the whole window. Anthropic has articulated this distinction in its engineering guidance, and Karpathy helped popularize the term. The rest of this article takes the emerging discipline apart. ...

A small model loop becomes a system only after we decide what it may remember, touch, spend, and survive. An agent demo can be a loop around an API call. A production agent is the surrounding discipline: state, tools, permissions, recovery, evaluation, and cost. This article offers a map of that discipline rather than a framework tutorial. Each pillar is described through the gap it fills, a minimal implementation, and the boundary where it stops helping. ...

A philosophical boundary and an engineering practice for continuity in AI agents Start by Making “Identity” Small Enough to Engineer Agent amnesia first damages the cost of cooperation, not the illusion of personhood. A strong session does not guarantee that the next one will preserve the same judgment. Users repeat preferences, teams restate constraints, and the system struggles to explain where an earlier decision came from. Long-term trust rests on these unglamorous forms of continuity: what the agent remembers, what it forgets, why it changed, and who approved the change. ...

This project note is part of my attempt to understand open-source AI systems by building with them, reading their migrations, and writing down where the abstraction holds—and where it leaks. Project learning list The problem is not remembering more An LLM can carry a conversation while the relevant messages still fit inside its context window. That is short-term continuity, not memory in the human sense and not durable application state. When the session ends, the model does not naturally retain that a user prefers terse answers, changed jobs last month, or abandoned an earlier plan. ...

Why UFO matters Desktop automation has always had an awkward boundary. Scripts are fast and precise when an application exposes the right interface. GUI automation is broad, but a changed label, a delayed window, or a custom-drawn control can break an otherwise sensible workflow. A vision-language model can interpret what it sees, yet perception alone does not make an action reliable. Microsoft’s UFO project is interesting because its evolution follows that boundary instead of pretending it does not exist. ...

LangChain used to be introduced as a box of chains, memory classes, prompt templates, loaders, and integrations. That description is historically accurate and operationally unhelpful. As of July 2026, the useful way to understand LangChain 1.x is much narrower: Call a model directly when the task is one inference with a clear input and output. Use create_agent when a model must choose among tools in a conventional model–tool loop. Use LangGraph when your application has durable state, explicit branches, retries, parallel work, long-running steps, or human decisions that are part of the workflow. This is not merely a taxonomy. It is an engineering rule about where complexity should live. Every framework abstraction creates a second system beside your product: another execution model, another vocabulary, and another place for failure to hide. The abstraction earns its place only when it makes the real system easier to operate. ...

A framework is most dangerous not when it lacks features, but when its old tutorials still look plausible. Familiar code can cross a version boundary long before it produces an obvious error. This guide follows the current LangChain Python 1.x path: model calls, agents, tools, structured output, middleware, memory, RAG, SQL, Ollama, and LangSmith. It is deliberately different from a framework overview. The aim is to leave you with a compact application architecture, runnable examples, and a way to recognize code that belongs to the 0.x era. ...

This article began as notes from a large-language-model meetup in January 2024. Back then, the question was what a model could do. Two years later, the harder question is why anyone should trust the system around it. I have not polished the old notes into a story in which every early judgment was right. Mistakes are useful sediment. They remind us that a technical opinion is not a prophecy; it is a bet with an expiry date. Each section therefore keeps the 2024 observation and adds a 2026 reassessment: still true, changed, or wrong at the time. ...

Status note, verified July 31, 2026: this article originally explained how to install the 2023 stand-alone Auto-GPT agent. Those commands are obsolete. The official project now says that AutoGPT Classic is unsupported, its dependencies will not be updated, and it has known security issues. Treat Classic as a historical laboratory, not a production tool. For new work, use the maintained AutoGPT Platform or another actively maintained workflow system. In the spring of 2023, Auto-GPT made a compelling promise: describe an objective, give a model some tools, and watch it plan its own way forward. I installed it because the idea felt less like a chatbot feature and more like a change in the shape of software. ...