Long-form notes on the products I have built and how things work in AI agent systems.https://benmcloughlin.com/en-gbhttps://benmcloughlin.com/learning/tool-error-handling-implementation/https://benmcloughlin.com/learning/tool-error-handling-implementation/The four-bucket mental model turned into code. Classifying errors by exception type and status code instead of trusting the message string; retrying transient failures in the harness with backoff and jitter; ending a run in code when there's nothing left to try; and a per-tool circuit breaker that stops a looping model without needing its permission - then all four layers assembled into one ordinary ReAct loop.Mon, 04 May 2026 00:00:00 GMTField noteagentstool-useerror-handlinghttps://benmcloughlin.com/learning/tool-error-handling/https://benmcloughlin.com/learning/tool-error-handling/The ReAct loop assumes the tool call returns something useful. In real work it often doesn't - the network times out, a credential expires, the thing you asked for isn't there. A walk through treating errors as just another observation, sorting every tool failure into four buckets, and the single rule that decides where each one goes: the harness retries what a retry can fix, the model sees only the errors its reasoning can move, and the rest end the run instead of spinning it.Sun, 03 May 2026 00:00:00 GMTField noteagentstool-useerror-handlinghttps://benmcloughlin.com/learning/parallel-tool-calling/https://benmcloughlin.com/learning/parallel-tool-calling/Most agents need to fetch a handful of independent bits of information before they can answer. Doing those one after another wastes most of the time. A walk through parallel tool calling - what it is, why the model decides when to use it, how it changes the shape of a ReAct turn, and how little code you have to write to get it.Sun, 26 Apr 2026 00:00:00 GMTField noteagentstool-usefundamentalshttps://benmcloughlin.com/learning/react-loop/https://benmcloughlin.com/learning/react-loop/ReAct - reason, then act, then observe, then reason again - is the loop running underneath every modern agent framework. A walk through what each step actually does, why mixing reasoning and acting fixed problems neither could fix alone, and the failure modes that hit production agents in the wild.Sun, 19 Apr 2026 00:00:00 GMTField noteagentsreactfundamentalshttps://benmcloughlin.com/learning/chain-of-thought/https://benmcloughlin.com/learning/chain-of-thought/Chain of thought started as a one-line prompt trick and ended up as the mechanism inside every reasoning model. A walk through what the trick was, why it worked, and the architectural decisions it forces on anyone building with LLMs in 2026.Tue, 14 Apr 2026 00:00:00 GMTField notereasoningchain-of-thoughtfundamentalshttps://benmcloughlin.com/products/data-scientist-agent/https://benmcloughlin.com/products/data-scientist-agent/An agent that takes on the working brief of an analyst or data scientist over a user's own dataset - fielding questions, running the right cuts, generating plots, and writing up reports. Full write-up coming once the build is far enough along to be honest about the trade-offs.Fri, 01 Nov 2024 00:00:00 GMTProductTypeScriptPythonAnthropichttps://benmcloughlin.com/products/data-viz/https://benmcloughlin.com/products/data-viz/A tool that picks the visualisation that actually fits the shape of the data, instead of asking you to choose from a gallery and hoping for the best. Full write-up coming once the build is far enough along to be honest about the trade-offs.Mon, 01 Jan 2024 00:00:00 GMTProductTypeScriptPythonAnthropic