标签：#AI

Introduction Speculative decoding is one of those techniques that has been "almost ready for production" for the better part of three years. A small draft model proposes tokens; a larger target model verifies them in a single forward pass. In theory, you get 2–4× throughput. In practice, the draft model has to be cheap, fast, and good enough at mimicking the target's distribution, which is a much harder combination than it sounds. Yesterday, a new paper from DeepSeek quietly climbed to the top of Hacker News (714+ points, 290+ comments at the time of writing). It's called DSpark , and it reframes speculative decoding in a way that looks like it could finally make the technique drop-in rather than bolt-on. The paper is here: github.com/deepseek-ai/DeepSpec/blob/main/DSpark_paper.pdf The Core Idea Instead of training a separate, smaller draft model from scratch (the classic approach), DSpark grafts the speculative head directly onto the target model. The intuition is simple: if the target model already knows which tokens are likely to follow, why not reuse its own intermediate representations rather than maintaining a parallel network? From the discussion on HN, this approach has a concrete architectural benefit — it reduces layer duplication that you'd otherwise have to maintain with a standalone draft model. In the DeepSeek experiments, the technique was applied on top of Step and Qwen 3.6 , which are themselves MTP-capable. How It Fits With MTP One of the more interesting practical points raised by HN commenters: DSpark is complementary to Multi-Token Prediction (MTP) , not a replacement for it. MTP — where the model predicts several future tokens at every step using auxiliary heads — has already been shown to give 50–100% speedups on hardware like the NVIDIA DGX Spark. DSpark adds another layer on top: even with MTP, the validation step is still a single forward pass through the main model, and the speculative tokens that get accepted come "for free." A useful men

Most job search tools focus on submitting more applications. I wanted to solve a different problem: reaching the people actually making hiring decisions. So I built PitchHired , an AI-powered platform that helps job seekers find hiring managers, generate personalized outreach emails, review them with AI, and send them from their own Gmail account on a business-hours schedule. The goal isn't to replace the job search, it's to remove repetitive work while keeping the candidate in control. I also chose a one-time credit model instead of monthly subscriptions because job seekers shouldn't have to keep paying while they're between opportunities. PitchHired is still evolving, and I'd genuinely appreciate feedback from fellow developers. What features would you want in a tool like this, and what would make you trust (or not trust) AI-assisted job search?

Headline: The wins come from context, not cleverness. An AI with your codebase, your design system, and your deploy logs in scope writes code that ships. Without that scope, it writes plausible code that doesn't. Two years ago, AI coding tools were autocomplete with attitude. In 2026 they are a credible second engineer — provided you build the workflow around them. This is the workflow I run today at Devya Solutions and on personal projects like eng-ahmed.com . The Stack Claude Code in the terminal — long-horizon, multi-file edits with skills and subagents. MCP (Model Context Protocol) servers for live access to docs, deployments, browser, and design tools. Cursor or VS Code for inline edits when I want to stay in the IDE. Why Context Is Everything The single highest-leverage move in AI-assisted dev is feeding the model the right context. MCP servers do this without prompt stuffing. Docs MCP — pulls current library docs at call time, so the model doesn't hallucinate the Tailwind v3 API in a v4 codebase. Browser MCP (Claude-in-Chrome) — lets the agent open the running dev server, screenshot the page, and verify the change actually rendered. Vercel MCP — fetches deploy logs and runtime errors directly. No more pasting logs. Context-mode MCP — keeps file scans, search results, and command output in a sandbox, only surfacing what's relevant to your conversation. A Real Workflow The blog page redesign I just shipped was built in a single 45-minute session. Rough flow: State the goal — two sentences, not a spec doc. Let the agent scout — Claude Code greps, reads a few files, proposes a plan. Iterate visually — screenshot the result, feed it back. The agent fixes the sticky-filter scroll bug in one turn. Commit and push — a single cm shortcut runs build, commits, and pushes. Vercel deploys on push. What the Agent Is Still Bad At Holistic taste — it copies the closest example in your codebase. If that's mediocre, the new feature is mediocre. Domain knowledge — it doesn't kn

What This Guide Covers When you start building LLM-powered applications, one pattern becomes unavoidable: RAG (Retrieval-Augmented Generation) . LLMs only know what they were trained on. Your company's internal documents, the latest spec sheets, project-specific information — none of that exists in the model. To handle data the model doesn't know, you need a system that retrieves relevant knowledge in real time and injects it into the context. That's RAG. In this guide, we'll implement a RAG system from scratch using pgvector and Gemini, then extend it step by step through Tool Use, AI Agents, MCP, and cloud deployment. Step 1: Embedding · Vector DB · RAG — core implementation Step 2: AI Architect perspective — design decisions explained Step 3: Tool Use — LLM autonomously searches the DB Step 4: AI Agents — combining multiple tools Step 5: MCP — exposing tools as a server Step 6: Cloud deployment — Render × Supabase Three Concepts to Understand First Embedding Computers can't measure "semantic similarity" from raw text. Embedding converts text into a list of numbers (a vector), and semantically similar words produce numerically similar patterns. "dog" → [ 0.82 , 0.75 , 0.10 , ... ] 768 numbers "cat" → [ 0.78 , 0.72 , 0.12 , ... ] ← similar pattern to "dog" "bank" → [ 0.08 , 0.10 , 0.85 , ... ] ← completely different Gemini's embedding model handles this conversion. Vector DB A regular DB searches by keyword matching. A vector DB searches by numeric distance — meaning it finds semantically related documents even when the exact words don't match. -- Regular search (misses if keywords don't match) SELECT * FROM docs WHERE body LIKE '%F1 score%' ; -- Vector search (finds semantically related docs) SELECT * FROM docs ORDER BY embedding <=> query_vector LIMIT 3 ; Search for "how to measure model performance" and it finds "F1 score calculation" — even without matching words. We use pgvector , a PostgreSQL extension, for this. RAG LLMs are limited to their training data. R

The Agent Skills open standard today, and the 2026 research on agents that write their own skills. TL;DR: In late 2025, "Agent Skills" became a thing — a dead-simple way to teach an AI agent a task: a folder with a SKILL.md file (some instructions in Markdown). It's already an open standard. The wild part is what's coming next: agents that write their own skills. I built a demo where an agent solves a task the hard way once, saves a real SKILL.md , and then reuses it — cutting its total effort almost in half. ~130 lines, no API key. First, what's a "skill"? If you've used Claude Code or similar tools lately, you've probably seen SKILL.md files. The idea is refreshingly low-tech. A "skill" is just a folder with a Markdown file that says how to do something : --- name : csv-to-markdown description : Turn comma-separated text into a Markdown table. Use when the input looks like CSV and the user wants a table. --- # CSV to Markdown ## Instructions Split the text into rows on newlines and columns on commas. Make the first row the header, add a `---` divider row, then format every row as `| a | b | c |`. That's it. No SDK, no config. Anthropic introduced this in October 2025 and then published it as an open standard ( agentskills.io ) in December 2025, so the same skill folder now works across ~30+ different agent tools (Claude Code, Cursor, Copilot, and more). The full rules are short ( agentskills.io/specification ): the only required fields are name (1–64 chars, lowercase-with-hyphens, and it must match the folder name) and description (≤1024 chars, saying what it does and when to use it ). Everything else — license , metadata , compatibility , allowed-tools — is optional. That's the whole spec. The SKILL.md files my demo writes follow it to the letter, so they'd load unmodified in any compatible CLI. The clever trick: progressive disclosure Here's the smart part. If you just dumped 50 skills' worth of instructions into the agent's context, you'd fill it up and leave n

Have you noticed how every demo of "an AI agent" looks impressive in the video and falls apart the moment you ask a sharper question? The agent confidently does the wrong thing. It forgets what it just decided. It tries to call a tool that doesn't exist. It loops forever rewriting the same file. It calmly tells you the deployment succeeded when it didn't. These aren't failures of the model. They're failures of the workflow around the model. Because that's all an agent really is: a software workflow where a language model can pick the next step and call tools. The "intelligence" sits in the prompt and the orchestration around it, not in some secret agent-flavoured fairy dust. Strip the word "agent" away and you've got five pieces of plumbing: planning, tool use, memory, constraints, verification. Every production-grade agent stands or falls on those five. This is a long walk through each one. Not the marketing version. The kind of detail you actually need before you ship something that talks to your database. The Loop You're Actually Building Before we touch any pillar individually, hold the whole loop in your head. A useful agent does roughly this on every turn: Read the goal (and whatever memory is relevant to it). Decide the next action: answer directly, call a tool, ask a clarifying question, or stop. If it called a tool, observe the tool's result and feed it back in. Update memory if anything is worth remembering. Check constraints: are we over budget, out of iterations, touching something off-limits? Verify the output before declaring success. Loop until done or stopped. That's it. Every framework (LangGraph, OpenAI Agents SDK, Claude Agent SDK, smolagents, whatever ships next month) is a different shape of the same loop with different defaults. agent-loop.ts async function runAgent ( goal : string , ctx : AgentContext ) { const state = ctx . startState ( goal ); for ( let step = 0 ; step < ctx . maxSteps ; step ++ ) { const decision = await ctx . model . decid

Active inference: curiosity emerges for free from minimizing surprise — 48% vs 100% on a foraging task. TL;DR: Most AI agents chase rewards — they pick whatever action scores the most points. I tried a different, brain-inspired goal: avoid surprises . Something neat happened — the agent became curious without being told to. It goes looking for information before acting, and that takes it from 48% to 100% on a simple task. ~100 lines. Two different ways to make decisions Most AI agents are "reward chasers." Give them points for doing well, and they'll pick whatever action they expect to score highest. Simple and effective. There's another idea from brain science: instead of chasing points, try to avoid being surprised — act so the world matches what you expected. It sounds almost too simple, but it leads to a surprising bonus: when you're trying not to be surprised, going and finding out what you don't know becomes valuable all by itself. In other words, curiosity isn't something you have to bolt on. It comes for free. This is called active inference , and in 2026 it jumped from neuroscience into AI as a serious approach ( here's a 2026 paper ). Here's the smallest demo that makes it click. The 10-second version The task: a reward is hidden behind either the LEFT door or the RIGHT door (50/50). There's also a hint you can check that tells you which door — if you bother to look. ❌ Reward-chaser ✅ Curious agent What it cares about getting the reward, right now getting the reward + not being unsure What it does guesses a door checks the hint first, then opens the right door Success (400 tries) 48% 100% Nobody told the second agent "go check the hint." It did it on its own, because being unsure bothered it. How it works Before acting, the agent scores each option on two things: Does this get me closer to the reward? Does this make me less unsure about what's going on? value_of_checking_the_hint = how_unsure_am_i # high when it's a total coin-flip value_of_just_guessing =

Theory of Mind and the Sally-Anne false-belief test, in ~60 lines of Python. TL;DR: There's a famous test that kids pass around age 4. It checks whether you understand that other people can believe things that aren't true. I built two AI agents: one that only knows "what's actually happening" (fails, like a toddler) and one that keeps track of what each person believes (passes). It's ~110 lines, and it's the foundation for agents that can actually work together . The test Sally puts her marble in the basket , then leaves the room. While she's gone, Anne moves the marble to the box . Sally comes back. Where will she look for her marble? If you said basket , nice — you just used something called "theory of mind." Sally never saw the marble move, so in her head it's still in the basket. What's actually true (it's in the box) and what Sally believes (it's in the basket) are two different things, and you kept them separate without even thinking about it. A 3-year-old says "box" — they can't yet separate what they know from what Sally knows. A 4-year-old says "basket." It's one of the most famous tests in child psychology, and in 2026 it's become a real test for AI agents too. The 10-second version ❌ Agent with no "theory of mind" ✅ Agent that models other minds What it tracks only what's actually true what each person believes, separately Where will Sally look? "box" "basket" Result FAIL (only knows reality) PASS How it works (the whole trick) The only difference between the two agents is one rule: a person's belief only updates when that person is actually in the room to see it happen. def someone_moves_the_marble ( new_place , who_is_watching ): for person in who_is_watching : # only people in the room beliefs [ person ] = new_place # update THEIR mental picture So when Anne moves the marble while Sally is out, only Anne's mental picture updates. Sally's is frozen at "basket." Ask the simple agent and it just reports reality ("box"). Ask the smarter agent and it answer

A sleep-like phase that consolidates noisy daily experience into durable memory — 75% vs 100% recall. TL;DR: There's a wave of 2026 research giving AI a "sleep" phase — time spent not answering questions, just tidying up what it learned that day. I built a 90-line demo of the idea. The agent that "sleeps" remembers 100% of what it learned. The exact same agent without sleep remembers only 75% and gets confused by bad info. Runs on a laptop. The memory problem every AI app hits If you've built anything with an LLM, you know the pain: the model only "remembers" what's in its current context window. Once the conversation gets long enough, the oldest stuff scrolls off the top and is just... gone. Forgotten. The usual fix is "make the context window bigger." But that's like fixing a messy desk by buying a bigger desk. It's expensive, and the model still gets worse as you cram more in (a real, measured effect — more text in the window can actually lower accuracy). Your brain doesn't work this way. You don't remember every sentence anyone said today. While you sleep, your brain replays the day, keeps the important bits as long-term memory, and dumps the rest. That's how you remember "I like coffee" without remembering every single cup. A couple of 2026 papers ask the obvious question: Do Language Models Need Sleep? Their answer: giving an AI a quiet "offline" phase to consolidate memories makes it remember better. So I built the simplest version that shows why. The 10-second version ❌ Agent with no sleep ✅ Agent that sleeps How it remembers keeps only the last N messages saves a tidy summary every night After 30 noisy days 75% recall 100% recall Tricked by bad info? yes no — it goes with what it saw most often Same experiences, same noise, same memory test. The only difference is whether the agent sleeps. How it works Each "day," the agent hears facts like Alice → drinks → coffee . To make it realistic, about 1 in 5 facts is wrong (people misremember, logs have errors). Th

A tiny Darwin Gödel Machine that edits itself and keeps only changes that verifiably score higher. TL;DR: I built a small program that improves itself . It looks at the tasks it's failing, edits its own code to fix them, and keeps a change only if the change actually makes it score better on a test. It goes from passing 1 of 8 tasks to 8 of 8 — and nobody wrote those fixes but the program itself. It runs on a laptop in under a second. No fancy hardware, no API key. The old dream: software that improves itself Normally, software only gets better when we make it better. You write code, you find a bug, you fix it, you ship again. The program never improves on its own. People have wanted "software that improves itself" for decades. The classic version (called a "Gödel Machine") had one rule that made it impossible to build: before the program could change a line of its own code, it had to mathematically prove the change would help. Proving that about real code is basically impossible, so the idea never worked. In 2025, researchers found a way around it with the Darwin Gödel Machine . They dropped the "prove it first" rule and replaced it with something every engineer already trusts: Try the change. Run the tests. If the score went up, keep it. If not, throw it away. That's it. It's basically how we all work — make an edit, run the test suite, keep what passes. The twist is that the program is the one making the edits. In the real paper, this let an AI coding assistant improve its own tooling and jump from solving 20% to 50% of a hard benchmark of real GitHub issues. I wanted to actually see this happen, so I built the tiniest version I could. The 10-second version Start After improving itself What it can do only uppercase learned 6 more skills on its own Test score 🔴 1 / 8 🟢 8 / 8 Who wrote the fixes? — the program did Start: ███░░░░░░░░░░░░░░░░░░░░░ 1/8 (only knows: uppercase) +reverse ██████░░░░░░░░░░░░ 2/8 +dedup_csv █████████░░░░░░░░░ 3/8 +sum_csv ████████████░░░░░░

I run an AI system that maintains itself on a schedule. One of its routines is supposed to do a job twice a week and save the result to a file. The scheduler swore it ran. Twice. lastRunAt right there - timestamped, green, smug. The file? Didn't exist. Not "saved in the wrong folder" - didn't exist anywhere. Here's the thing nobody warns you about when you wire up autonomous agents: "it ran" and "it worked" are different claims, and most of your dashboards only check the first one. The trap A scheduler firing a job tells you a process started . It tells you nothing about whether the job did the thing. My routine started, hit an early error reading a file that didn't exist yet, and just... ended. No crash. No red anywhere. It "ran." It produced nothing. For days. If I'd trusted the green checkmark, I'd still think it was fine. How I found it I stopped reading the status and went to the disk. Three checks, in order: Does the output actually exist? Not "did it run" - does the artifact it's supposed to produce exist, right now, where it claims to put it? If yes - is it fresh and non-empty? A stale or empty file is a silent failure wearing a costume. If no - read the raw run log. Not the summary. The actual transcript of what the agent did, tool call by tool call. That third check is where the truth was hiding. The summary said the routine was "episodic." The transcript said something blunter: it tried to read its own memory file, got "file does not exist," and never recovered to create it. Zero write calls the entire run. It never even tried to save anything. "Episodic" and "dies before it writes" lead to completely different fixes. The summary would've sent me down the wrong one. Steal these If you run anything autonomous: "Ran" is not "worked." Health is the artifact: it exists, it's fresh, it's not empty. Not a green dot from the thing that launched it. Described is not executed. What the spec says a routine does is a hypothesis. What's on disk is the fact. When they

Teenage Engineering has already issued multiple substantial updates for its surprisingly capable $329 EP-133 KO II sampler. Its latest is one of the biggest yet. OS 2.5 adds audio over USB, selectable sample rates for lo-fi fun, sample reverse, an arpeggiator, equal-length autochopping, and it extends the maximum length of a sample from 20 seconds […]

Series: AI, Ego & Regret — Bonus Chapter Editor's Note: While compiling the old series for the...

Not everyone is buying Elon Musk’s vision for orbital data centers.

Hello Devs 👋 AI coding assistants have changed the way many teams build software. Tasks like generating components, creating tests, writing boilerplate, or handling repetitive refactors can now happen in minutes instead of hours. The productivity gain is real and that part is easy to notice. What becomes interesting after using these tools for a while is that a different bottleneck starts appearing. Code generation becomes faster, but the review process often stays the same. Teams can generate hundreds of lines of code within minutes, but someone still has to answer important questions: Does this actually solve the requirement? Are edge cases covered? Will this introduce side effects? Does it align with existing patterns? The speed of writing code has changed. The need for confidence has not. That is where I think the conversation around AI-assisted development is starting to shift. The challenge is becoming less about generating code and more about making sure the generated code is actually safe to ship. The Problem With Reviewing AI-Generated Code Like Regular Code Imagine asking an AI coding assistant to implement coupon validation for premium users. Add coupon validation for premium users and create tests A few seconds later you get: if ( user . isPremium ){ applyCoupon (); } Nothing immediately looks wrong. The code is clean, there are no syntax issues, tests may pass, and the implementation appears complete. But pull request reviews usually go beyond reading diffs. Reviewers start asking questions such as: What happens if the coupon has expired? Does this affect payment calculations? Should audit logs be updated? Are there services depending on this behavior? This is where AI-generated code becomes interesting. It can often be functionally correct while still missing important implementation details. Research around larger AI-generated projects has also shown that functional correctness does not necessarily translate into maintainable system design. Teams stil

Maraget Atwood, the storied author of The Handmaid's Tale and The Blind Assassin, was interviewed as part of the Babell Literary and Cultural Festival in Porto, Portugal. As it usually does at these things, the issue of AI came up, and Atwood didn't mince words. According to Deadline's recap, Atwood said she'd used an AI […]

Google's NotebookLM is great. But handing your research notes, PDFs, and meeting transcripts to Google's cloud is a hard sell for a lot of people — especially when those documents contain client data, unpublished research, or internal strategy. So I spent a weekend testing five open-source alternatives. Three things mattered: can I docker compose up in under 10 minutes, does the podcast feature actually work offline, and what breaks first? Here's what I found. The Contenders Project Deploy Time Min VRAM True Offline License Open Notebook (lfnovo) ~8 min 8 GB Yes MIT Notex (smallnest) ~3 min 4 GB Yes Open source KnowNote (MrSibe) ~2 min 4 GB Yes Open source NotebookLM-Local (nagaforcloud) ~15 min 8 GB Qwen-3 4B bundled Open source InsightsLM (phsphd) ~30 min 8 GB Yes N8N SUS license 1. Open Notebook — The One to Beat git clone https://github.com/lfnovo/open-notebook cd open-notebook docker compose up -d Eight minutes from git clone to the web UI on localhost:3000 . It ships with 18+ model providers pre-configured — Ollama, OpenAI, Claude, DeepSeek, Gemini, all selectable per notebook. The podcast generator supports 1-4 speakers with different voices, and it runs entirely offline when you point it at an Ollama backend. What works: Document ingestion is fast — SurrealDB's vector + full-text index handles 200-page PDFs without choking Model switching is genuinely useful — Claude for deep analysis on one notebook, local Qwen for quick summaries on another Podcast quality with 2 speakers is close to NotebookLM's. 4 speakers is still rough. What breaks: Citation highlighting is still being rebuilt (work in progress as of June 2026) Single-user only — no team/workspace isolation built in Docker required. No native binary. 2. Notex — Single Binary, Zero Dependencies Notex is written in Go. You download a single binary (~25MB) and run ./notex . That's it. No Docker, no Python venv, no database setup. It supports PDF, TXT, MD, DOCX, HTML, audio, and YouTube/Bilibili URLs as so

Everyone hits the same wall with Claude Code. You add a rule to CLAUDE.md . It works. You add ten more. They mostly work. You add forty more — and now Claude is cheerfully ignoring the rule you care about most, the one that's been sitting there since day one. So you make it LOUDER , in caps, with three exclamation points. It still gets skipped. The instinct is to write more. The fix is almost always to write less . Here's why, and what to do instead. Instruction-following has a budget, and you're overdrawn This is the part most CLAUDE.md guides skip. Frontier models reliably follow on the order of 150–200 instructions at once — and adherence to any single rule drops as you stack more on top of it. It isn't a cliff; it's a slow tax. Every line you add makes every other line a little less likely to be honored. Now subtract what you don't control: Claude Code's own system prompt already spends a chunk of that budget before your file is even read. So the working budget for your project rules is smaller than the headline number — and a sprawling 300-line CLAUDE.md isn't 300 rules followed, it's maybe the first 150 followed well and the rest treated as ambience. The mental model that fixes everything downstream: CLAUDE.md is a budget, not a wishlist. You are not writing documentation. You are spending a scarce attention allowance, and every line competes with every other line. The test for every line: would you bet $5 it's followed? Go through your file line by line and ask one question of each rule: would I bet money this fires every time it's relevant? Three outcomes: Yes, and it's load-bearing — keep it. This is what the budget is for. Nice to have, but I wouldn't bet on it — cut it, or move it to a referenced file (below). It's diluting the rules you would bet on. It absolutely must happen every time — then it doesn't belong in CLAUDE.md at all. Make it a hook. That third category is the one people get wrong, so let's be concrete about it. Advisory vs. deterministic:

I've been building agent workflows for a while, and one capability keeps coming up that the ecosystem hasn't fully reckoned with: letting an AI agent publish a document to the public internet and hand someone a link. It sounds trivial ("save HTML, return a URL"). It isn't. The moment an autonomous agent can mint a public link, you've handed it a primitive that touches access control, data exposure, and reputation. This post is about the design questions that surface once you take that seriously, written by someone who builds in this space. Disclosure up front: I work on Thryvate, a document-sharing tool with an MCP server. More on that at the end, but the problems below are general. The naive version The first version everyone writes is a tool that takes content and dumps it to object storage behind a public CDN URL: publish(html) -> https://cdn.example.com/a8f3c2.html Ship that and an agent can now share its work. It can also now: expose a half-finished draft to anyone who guesses the URL, leave that URL live forever with no way to pull it back, publish something containing a customer's name with zero record of who saw it. For a human hitting "publish" deliberately, those are acceptable defaults. For an agent doing it as one step in a longer plan, they're landmines. What "publish" should actually mean for an agent A few properties turn the naive primitive into something you'd trust an agent to call: 1. Default to private, opt into public. The safe default for an agent-minted link is not "world-readable." It's "only people on this list" or "only people with the password." Public should be an explicit parameter someone has to set, not the fallback. 2. Revocability. Anything an agent publishes, you must be able to un-publish instantly. A live link is a liability with a half-life, and the ability to revoke is what makes it safe to let the agent create them liberally. 3. Expiry as a first-class field. "This link dies in 7 days" should be a parameter on the publish call,

Hello, I'm Maneshwar. I'm building git-lrc, a Micro AI code reviewer that runs on every commit. It is free and source-available on Github. Star git-lrc to help devs discover the project. Do give it a try and share your feedback. A single honeybee has exactly one move: find nectar, fly it home. Impressive aviation. Add a few thousand more bees and something strange happens. Now they're making honey, cooling the hive, and defending the colony against threats ten thousand times their size, with no Jira board, no standup, and nobody handing out tickets. That jump from "can fetch nectar" to "runs a self-regulating honey factory" is the best mental model I've found for multi-agent AI systems . So let's steal it xD First, what even is an "agent"? Before we throw thousands of them at a problem, it's worth pinning down what one actually is. An AI agent is an autonomous system that performs tasks on behalf of a user (or another system) by designing its own workflow and using available tools . Three things decide how good an agent actually is: The LLM powering it i.e the brain. Its tools which is the hands. The reasoning framework is how it turns tool outputs into the next decision. A single agent is fine. It's our lone bee, and it can do real work. But ask it to research a topic, run heavy calculations, scrape five websites, and write the summary, and you start to feel the ceiling. Multi-agent systems: bees, but for compute A multi-agent system keeps each agent autonomous but lets them cooperate and coordinate inside a structure . The magic isn't any single agent, it's the choreography between them (claude which is famous for that). And there are a few classic ways to choreograph it. 1. The decentralized network (a.k.a. "everyone's a peer") Every agent can talk to every other agent. They share information and resources, and they all operate with the same authority . No boss. Just message-passing. This is your agent network . It's great for emergent, collaborative problem-solv