What Are AI Agents? A Complete Guide
An AI agent is a language model placed inside a loop with tools and a stopping rule. How that loop works, where it breaks, and when not to build one.
A. Rahman10 min read
Loading the homepage
Explore the systems, tools, agents and technologies shaping the next generation of software.
The pieces we would hand a new reader first — the ones that set up everything else on the site.
An AI agent is a language model placed inside a loop with tools and a stopping rule. How that loop works, where it breaks, and when not to build one.
MCP replaces N times M model-to-system adapters with one server many clients can use. What the protocol standardises, and what it deliberately leaves to you.
Rule-based automation fails on inputs it was not written for. Agents fail unpredictably. A structured comparison of where each belongs, and how to combine them.
Build a tool-calling agent as a Next.js route handler: typed tool definitions, a bounded loop, a streamed step log, and a client leaf that renders progress.
RAG solves a knowledge problem; agents solve an action problem. Where each architecture belongs, how they combine, and why their failure modes barely overlap.
The specialised roles this publication covers — and, in the case of the SEO Agent, the one it runs on its own articles.
Analyzing technical knowledge
Reads primary sources — specifications, changelogs, papers and repositories — then reduces them to the handful of claims that actually matter, each traceable back to where it came from.
Reading and writing software
Works inside a repository rather than a chat window: reads the surrounding code, proposes a change, runs the test suite, and iterates on the failure output until the change actually holds.
Auditing search visibility
Audits metadata, heading hierarchy, structured data and internal links against a published scoring model, then returns ranked recommendations for a human to approve — it never rewrites content on its own.
Planning editorial coverage
Maps a topic into an outline, finds the gaps a subject already covered elsewhere leaves open, and proposes structure — the writing and the judgement stay with the author.
Interrogating structured data
Turns a question into a query, runs it, and reports both the answer and the shape of the data behind it — including the rows that do not fit the story.
Orchestrating workflows
Chains tools, APIs and human approval steps into a durable workflow, with retries and checkpoints so a failure halfway through does not mean starting over.
Filter by section to narrow the list. Nothing reloads — the whole archive is already on the page.
An AI agent is a language model placed inside a loop with tools and a stopping rule. How that loop works, where it breaks, and when not to build one.
Rule-based automation fails on inputs it was not written for. Agents fail unpredictably. A structured comparison of where each belongs, and how to combine them.
Build a tool-calling agent as a Next.js route handler: typed tool definitions, a bounded loop, a streamed step log, and a client leaf that renders progress.
RAG solves a knowledge problem; agents solve an action problem. Where each architecture belongs, how they combine, and why their failure modes barely overlap.
MCP replaces N times M model-to-system adapters with one server many clients can use. What the protocol standardises, and what it deliberately leaves to you.
A criteria-first assessment of AI coding tools by category — completion, repo chat, agentic editors and review bots — with a harness for testing them yourself.
A comparison of what actually differs between the two API surfaces — system prompts, tool schemas, structured output, state and streaming — and how to choose.
When a change can be drafted by a process that runs your tests, the constraint moves from writing code to reviewing it. What that shift breaks and rewards.
The layers between a working AI demo and a system people depend on: call boundaries, budgets, tenant isolation, evaluation, observability, degradation.
Every article lives in one section of the taxonomy, so a subject can be read end to end instead of in the order it happened to be published.
Agents
AI agents combine a language model with memory, tools and a control loop so software can pursue a goal rather than answer a single prompt. This section covers agent architectures, planning…
AI
Applied artificial intelligence for people who build software: how large language models actually work, what generative systems can and cannot do, where the research is heading, and how to…
A working reference rather than a leaderboard: what each tool is genuinely good at, and where it stops being the right answer.
Unified model interface for JavaScript applications
The AI SDK gives JavaScript and TypeScript applications a single interface across model providers, covering text generation, structured output, tool calling and embeddings. Its streaming…
Best forWeb applications that need streaming AI interfaces today and the freedom to swap model providers later.
Agentic coding in the terminal
Claude Code is Anthropic’s coding agent that runs as a terminal process and operates directly on a repository — it reads files, edits them, runs commands and iterates on the output it gets…
Best forMulti-file changes where the agent needs to read the surrounding code, run the test suite and iterate on real failure output.
AI-native code editor
Cursor is a fork of VS Code rebuilt around model-assisted editing: inline completion, a chat pane with codebase context, and an agent mode that applies edits across several files. It…
Best forDevelopers who want agentic edits and repo-aware chat without leaving a familiar VS Code environment.
Completions and chat inside every editor
GitHub Copilot provides inline completion, chat and agent-style edits across the major editors, backed by a selectable model. It draws context from open files and the surrounding project,…
Best forOrganisations already on GitHub that want one governed assistant available in every editor their teams use.
Open-source LLM tracing and evaluation
Langfuse captures traces of LLM applications as nested spans — one per model call, retrieval step and tool invocation — with inputs, outputs, latency and token counts attached. On top of…
Best forTeams that need production tracing and evaluation while keeping trace data inside their own infrastructure.
Graph-structured agent runtime
LangGraph models an agent as a state graph: nodes are steps, edges are transitions, and a shared state object flows between them. Because control flow is explicit rather than implicit in a…
Best forLong-running or approval-gated agent workflows that need durable state and control flow you can read off a diagram.
A technology publication for the people building on top of language models — written at the level where the decisions actually get made.
Start here
Tech Agents is a technology publication about the software layer that has grown up around large language models — the agents, protocols, frameworks and infrastructure that turn a model into something a team can ship and then keep running. It is written for the engineers, technical leads and founders who have to choose an architecture and then live with it. It exists because most writing about AI lands at one of two extremes: launch summaries that explain nothing, or research papers that assume you already have the background. Engineering decisions get made in the gap between those two, and that gap is the brief.
The practical questions tend to arrive in the same order. What is an agent, and how is it actually different from a workflow with a model call in the middle of it? How do you give one tools without handing an attacker a privileged API? When does retrieval answer your problem, and when are you really describing an agent? How do you stream a generated response through a React tree so that a cancelled request tears down cleanly instead of leaving half an interface behind?
The article archive is organised around questions like those. Tutorials are written to be run rather than admired: the code in them is the code that made the thing work, including the error handling that usually gets edited out for brevity. Explainers start from a concrete failure — a retry loop that swallows the one error worth seeing, a retrieval query that quietly returns another tenant’s rows, a tool schema the model keeps filling in wrong — and work outward to the principle underneath it.
Four strands run through the site. AI agents covers architecture: planning loops, tool calling, memory and context budgets, multi-agent orchestration, and what breaks the first time a loop runs unattended. AI development is the engineering discipline around shipping those systems — retrieval pipelines, streaming interfaces, evaluation harnesses, and cost and latency budgets treated as design constraints rather than afterthoughts. Security treats prompt injection and tool permissions as an architecture problem rather than a filtering problem, because that is where the fix has to live.
The fourth strand is tooling. The AI tools directory is a working reference to coding assistants, agent frameworks, vector stores, orchestration platforms and model providers, each described by what it is genuinely good at and where it stops being the right answer — deliberately without scores, because a single number hides the trade-off you came to read about. Around all of it sits the reference layer: a glossary for the vocabulary that gets used inconsistently everywhere else, and side-by-side comparisons for the decisions that keep coming up in the same shape.
Two rules shape what gets published. The first is that facts are never invented to fill a gap. Benchmark scores, adoption figures, funding rounds, star counts and quotes are the easiest things in the world to fabricate and the hardest for a reader to check, so where a number cannot be verified you will find the mechanism instead — the variables that move the result, and the trade-off you are actually choosing between.
The second is that primary sources beat commentary. Specifications, vendor documentation and changelogs are linked directly from the references at the end of an article, and where a claim depends on a version, the version is named. Every piece carries a publication date and, once it has been revised, an updated date, because writing about a fast-moving protocol is only as useful as its timestamp. Every article also names the desk it came from — and the four contributor profiles behind those bylines are placeholders rather than real people today, which the about page states plainly rather than leaving you to work it out. If something here is wrong, out of date, or missing the caveat that would have changed your decision, the editors would rather hear about it.
Get the latest AI, developer tools and emerging technology insights — without the noise.
One issue every Tuesday. Unsubscribe in one click. Privacy policy.