Folarin Akinloye is an AI Engineer based in London, UK. He builds production-ready agentic AI systems, multi-agent architectures, and sophisticated RAG implementations, and writes about the engineering decisions behind them.
A skill bundles two attack surfaces in one folder: a body that becomes trusted context the moment it triggers, and scripts that run with real privileges. Add supply-chain risk and you have a dependency, not a document.
A tool is one capability, MCP is connectivity to external systems, a Skill is packaged know-how for using what you already have, and a subagent is an isolated context window. How to pick the right layer instead of the wrong one out of habit.
An Agent Skill packages one piece of procedural expertise into a folder the model loads on demand. Progressive disclosure is what makes it cheap: install a hundred skills and pay for the one that fires.
The design job for an AI feature is to capture the model's value while making its uncertainty legible and its errors survivable. Stream for perceived speed, cite for trust, treat the unhappy path as the main work, and capture feedback to fuel the flywheel.
A production AI app is a pipeline of checkpoints around one model call: guard the input, pick the simplest orchestration that works, route through a gateway, guard the output, and always have a deterministic fallback. Here is the whole map.
The hard part of an AI product is not the code, it is the judgment about what "good" means. That needs a single owner, not a committee, and an incident response loop that ends in a new eval row.
Hallucination, injection, and drift are properties of the technology, not bugs you fix once. The senior move is to detect, contain, and monitor every one of them, and to keep a live risk register.
SaaS has near-zero marginal cost. An AI product pays real inference money on every request, so you can ship something users love that loses money per use. Here is how to model and control that.
An AI product silently degrades weeks after launch with no code change. LLMOps is the loop that keeps it good, and the data flywheel is how production usage compounds into a product a competitor cannot clone.
An AI feature will be wrong some fraction of the time, forever. So do not gate ship on "no bad outputs." Gate it on containment: narrow scope, a staged ramp with checkpoints, a deterministic fallback, and an instant undo.
The model is a third-party dependency that reprices, deprecates, and changes behaviour on a schedule you do not control. None of that is a bug in your code. Here is how to build so it does not surprise you.
LLMs fit language work that tolerates some imperfection and is expensive to do by hand. Here is a concrete test for whether your problem qualifies, and what to do when it fails the test.
Writing your evaluators up front, like TDD for prompts, backfires. Evals have to be discovered from real failures. Here is the discipline that actually works, and how small a real suite ends up being.
Model capability is a commodity you rent by the token. If a competitor's model gets twice as smart next quarter, do you die or thrive? Here are the moats that survive, and the ones that only feel like moats.
A working AI demo takes an afternoon. A product a business runs revenue through takes a quarter. The gap is not model quality; it is the 95% wrapped around the model. Here is what that 95% is.
A hands-on guide to running Phoenix locally: trace an agent over OpenTelemetry, then score it with LLM-as-judge and code evaluators.
AgentOps is the practice of running agents in production: observability, evaluation, guardrails, cost governance, and human oversight.
OpenTelemetry in plain terms: the data model, the GenAI semantic conventions, and why building agent tracing on OTel avoids lock-in.
What traces and spans actually are, how they map onto an agent run, and how to walk one to find the exact step that broke.
A tour of the reasoning research: how prompting elicits it, how researchers categorize the techniques, and why one prominent position paper says it is not reasoning at all.
Most prompting advice transfers across models. The parts that do not will quietly break your app when you switch providers. A working map of the differences.
A production case study measured fourteen prompt modifications one by one. The results are humbling: naming the model helped, few-shot examples hurt, and templates cost accuracy.
Prompt functions wrap a prompt in a name, inputs, and a rule so you can call and compose it like code. The chat-window version is dated; the idea runs everything now.
Code generation is the most-used LLM capability and the most casually prompted. A tour of the core patterns, plus the checks that catch what the model quietly gets wrong.
Cranking temperature does not make synthetic data diverse. Injecting randomness into the prompt does. Here is the technique, with code.
Most agents fail for boring, fixable reasons. Here are the practices that make an agent reliable: fewer LLM calls, tools that explain themselves, and a debugging ladder to climb when it misbehaves.
When retrieval fails in your domain or language, you can prompt a large model to generate training data for a custom retriever, matching thousands of hand-labels for tens of dollars.
You can prompt a model to produce labeled examples on demand: test fixtures, eval sets, cold-start training data. Here is how to do it well and where the sharp edges are.
Your few-shot examples carry hidden biases. The distribution and order of them can nudge the model toward a label before it reads the input. Here is how to spot it and fix it.
Models hallucinate because they are built to sound coherent, not to be right. Here are three prompt-level moves that reduce made-up answers, plus where prompting stops working.
Prompt injection, prompt leaking, and jailbreaking are three different attacks with three different fixes. Here is how each one works and what defence actually holds.
Graph prompting feeds a model structure, not just prose: entities as nodes, relationships as edges. Here is what the idea is, where it comes from, and when it pays off.
smolagents ships two agent classes that act in completely different ways: one writes Python, one emits structured JSON. Here is how each works, what each is good and bad at, and when to reach for which.
Reasoning models do the chain-of-thought for you. That flips a lot of prompting advice on its head, and getting it wrong wastes tokens and quality.
Multimodal CoT extends step-by-step reasoning to text plus vision with a two-stage split: build a rationale from both modalities, then answer from the rationale.
Chain-of-Thought lets a model reason and compute in the same breath, which is where it slips. PAL keeps the reasoning in language but hands the math to a Python interpreter.
Directional Stimulus Prompting trains a small model to produce hints that nudge a big frozen LLM toward what you want. It is prompt optimization when you cannot touch the weights.
Hand-picking chain-of-thought examples is a guess. Active-Prompt uses model disagreement to find the questions where a human annotation actually moves the needle.
ART made a frozen LLM generate its own multi-step reasoning, pause to call tools, and resume. It is 2023 research, but the pattern is everywhere in modern agents.
For questions that need world knowledge, having the model generate relevant facts first, then answer using them, beats answering cold. How it works, and when to use it over RAG.
One chain of thought can take a wrong turn. Self-consistency samples several reasoning paths for the same question and votes on the answer, beating a single greedy pass on reasoning tasks.
Telling a model to reason before it answers genuinely improves accuracy on multi-step problems. How standard and zero-shot chain-of-thought work, when to use them, and the cost.
Examples are the most effective context you can add to a prompt, until they aren't. When few-shot beats zero-shot, how many to use, how to pick them, and where examples won't help.
A good prompt has up to four parts: instruction, context, input data, and output indicator. Here is what each does and the design habits that make them land.
Before you rewrite a prompt, check the dials around it. Temperature, top_p, max tokens, and the penalties each change your output in a specific way. Here is when to touch each.
An agent is a stateless model wrapped in a loop. Here is the ReAct pattern built from scratch: the message list as memory, the exact algorithm, and the rule that generate calls equal tool rounds plus one.
Git worktrees let you check out several branches into separate folders at once. They went from obscure to default because running coding agents in parallel needs isolated workspaces.
Anthropic's Claude Tag drops a shared, persistent agent into Slack channels. The interesting part is not the model, it is the move from one-off chats to a teammate that remembers.
Why model deprecation quietly breaks production apps, and a concrete workflow: one models constant file, a scheduled GitHub Action checking deprecations.info plus a liveness probe, and an alert.
A practical guide to LLM observability: the traces and metrics to log, the handful of alerts worth waking up for, and why traditional monitoring misses what matters.
How to choose how to structure an agent: the plain while loop, the explicit state machine, and the full graph. What each gives you, what each costs, and when to move up.
Prompt caching is solved. The bigger win for agents is caching tool-call results, and that is harder because tools have side effects. Here is how to do it safely.
How to build a RAG system that serves many customers without leaking data between them: silo vs pool vs bridge, namespaces vs metadata filters, and where isolation must hold.
How to stream LLM tokens, tool progress, and state updates from a LangGraph agent using the v2 stream API, with the custom writer pattern most people miss.
If you can already ship software, the way into AI engineering is not another tutorial. It is building five systems that force you to handle failure, measure quality, and run things in production.
Autonomy is great until the agent deletes the wrong file or emails the wrong list. Here is how DeepAgents pauses on sensitive tool calls and waits for a human, in TypeScript.
Skills give an agent deep, reusable capability without paying for it on every turn. Here is how DeepAgents loads them in layers, and how to write ones that get picked.
DeepAgents treats long-term memory as files the agent reads and writes, with backends deciding where they live. Here is how scoping, episodic memory, and consolidation actually work.
Subagents are the main lever for stopping a long task from drowning in its own context. Here is how DeepAgents implements them in TypeScript, and the mistakes that make them useless.
A tour of how DeepAgents assembles, compresses, and isolates context, with the TypeScript config for each layer. The framework does a lot automatically; this is what is happening and where you steer it.
Prompt caching can cut input cost by up to 90%, but only if you structure prompts so the cache actually hits. Here is how it works and when it pays off.
What state, checkpoints, and threads actually are in LangGraph, how to pick a checkpointer, and how persistence powers memory, resume, and human-in-the-loop.
A practical playbook for making LLM apps cheaper and faster without hurting answers: where the cost and latency actually go, and the levers that work.
A practical guide to keeping production agents safe: where to put input and output rails, how to contain tools, and why prompt injection needs layers, not a fix.
A full walkthrough of streaming LLM output from the provider, through your backend, into a React UI: SSE, backpressure, cancellation, and the gotchas.
The field moves faster than any course can. So I learn production AI by gutting and rebuilding the open-source projects that production teams actually run. Here is the path.
A deep, hands-on tour of Vercel's Eve framework for durable agents: the filesystem-first model, every core capability, a complete TypeScript research agent, and an honest comparison with LangChain's Deep Agents.
A clear split between an agent's working memory and its cross-session memory, with LangGraph code for checkpointers, stores, and the semantic/episodic/procedural breakdown.
Structured outputs guarantee the model's response matches your JSON schema. Here is how it works under the hood, how it relates to function calling, and the patterns that hold up in production.
A practical, opinionated comparison of the four vector databases people actually argue about, with a clear default and the signals that mean you have outgrown it.
Embeddings turn text into vectors so meaning becomes geometry. Here is the mental model, the gotchas, and how to choose a model in 2026.
Rerankers re-score your top candidates with a model that actually reads the query and the document together. Here is how they work and when they pay off.
Chunking quietly decides how good your RAG system can ever be. Here is how the main strategies compare and a simple way to pick one.
Deep Agents bakes planning, context management, subagents, and human approval into the agent loop. Here is why it exists, every main feature, and a production-grade customer support agent built with it.
You cannot improve an agent you cannot measure. A deep, practical guide to evaluating agents with LangSmith: tracing, datasets, evaluators, experiments, and what makes agent evaluation different.
A hands-on guide to building MCP servers with FastMCP, drawn from two I actually shipped: the anatomy, the rules, stdio vs remote HTTP, the pitfalls, and how to test and deploy.
An agent with no tools is just a chatbot. Here is how tool calling really works, what makes a tool the model can use reliably, and how MCP lets you plug in tools you did not write.
Traditional RAG retrieves once and hopes for the best. Agentic RAG lets the model decide whether, what, and how to retrieve, then check its own work. Here is the difference, with code.
Agents, explained without the hype: what they are, when one is enough, when you need several, and how to build a working multi-agent system with LangChain.