We’re setting out to build an AgentA system that wraps an LLM with tools, memory, and a loop, so it can take multi-step actions toward a goal rather than just answering one prompt. that plans a family road trip, the days and the stops inside them, with everyone in the family adding to the plan. It runs on AWS, and Coding agentAn agent whose tools are the ones a programmer uses, so it can read a codebase, change it, run the tests and keep going until the task is done or it gives up. write it. Once it works, we take it on a real trip and write that up. Every example follows one trip: Perth to Broome and back in the July 2027 school holidays, for a family whose wants pull in different directions, so no plan can give everyone everything.
We make every design choice for this family, their budget and the roads of remote WA. The Exam Room covers most of the same ground in general terms, from running agents in production to scoring an agent’s run as well as its answer. Here those posts are background, not authority. Where our answer differs from theirs, I say why.
If you build software and want to watch an agent go from an idea to a printed itinerary without the marketing, this is written for you. It’s also for anyone planning a drive through remote Australia, and for any parent who has refereed the back-seat argument about whose turn it is. You’ll see every choice next to the ones we passed over, and whenever I say the agent has got better, I’ll show you how I know.
The posts show the coding agents’ working as well as their results. Each post that builds something ends with a section on how it was built. It covers what the coding agent was asked, what came back, and what review and testing caught before anything merged. The coding agents work from a written brief kept in the repository with the code. The Greenbox team found out why that brief is needed when they started teaching their LLM the codebase: two developers asked the same model for the same function and got two different functions.
Several people plan one trip. Every change to the plan records who asked for it, so a day that moved traces back to whoever moved it. The agent then reports how fairly the finished plan serves each person. That report is the part I most want, and the part I’m least sure how to build.
Map coverage in remote WA is thin between towns. The closures that decide whether a highway is open on the day come from the state. A plan built from the map alone can route a family down a highway the state closed that morning.
The first question is whether this needs an agent at all. Much of a trip plan is arithmetic and lookup: distances, fuel ranges, opening hours, whether a day fits before dark. That’s ground where an LLM is the wrong tool, and an agent is an LLMA neural network trained to predict the next token in a sequence, large enough that it generalises to tasks it wasn’t explicitly trained for. with Tool useLetting an LLM call structured functions you’ve defined – search, calculator, database query, API call – instead of trying to do everything in text.. A routing API and some ordinary code might do the whole job. So we start there, with the case for and against, before anyone writes a line.