Becoming Engineers
In practice
Barnaby Robson
Act one
The harness
The setup that puts the model to work.
An agent = its model + its harness
Agent
Model
Harness
The model interprets the task and proposes actions. The harness supplies context and tools, checks permissions, and runs those actions. That loop is what lets an agent work on your files.
NextThe device. It hosts the tools and files.
Why use the command line
What it gives an agent
I use a computer because the agent can work directly with my files and installed tools. A server can provide the same working environment, with a phone or tablet as the interface.
NextThe interface. The terminal has been there the whole time.
Choose where you run the agent
Built in
Run an agent CLI that you install separately.
Terminal workspaces
Run separately installed agent CLIs, side by side.
Editors
Integrated agents; supported models and extensions vary.
Provider apps
The provider packages the agent and its interface.
The first two columns host the agent CLI you install. The other two integrate agent features into an application. I use terminal workspaces to keep several providers’ CLIs together.
Nextcmux. Three words carry every later screen.
Workspace, pane, surface
Three words
You do not have to be a developer to open this. It starts like any other application, and the workspace holds one piece of work the way a folder does. A surface can be a browser, so the agent can open a page for you.
NextWhat the leaderboards on that page tell you.
Different models, different jobs
Read the benchmark and its method
Speed, quality and cost are separate measures. I use the boards to shortlist models, then check them on my own work. Subscription allowance is a separate limit from API cost.
NextGive each agent a short command.
Give each agent an alias
Four agents, one way of working
An alias is a short name for a command. These are my local shortcuts. Their flags bypass approval checks; the model comes from the tool’s current configuration.
NextYou can talk to it.
Say the prompt
“Read the meeting note, list the open questions, and link each one to its source heading.”
I use dictation for longer requests. Read the transcription before sending it: a misheard file name can send the agent to the wrong place.
NextAct two. What any one of them knows.
Act two
Context engineering
Choose the information the agent needs now.
What goes into the first answer
You launch it
It assembles context
You give it a job
It retrieves more
Before answering, the model receives instructions and context from the runtime. I shape both the standing context and the evidence for this job. A file path tells it where to look; the file still has to be read.
NextEverything loaded shares one context window.
Smart zone, dumb zone
One context window
Instructions · prompts · file contents · tool results · conversation
A focused task, with room to work.
Illustrative boundary. Behaviour varies by model and task.
Everything loaded competes for one window. As it fills, recall and task performance can fall. At a new phase, I use handoff to carry the goal, key evidence and next step into a clean session.
NextMemory is one source of that context.
Memory — what does it know about you?
“Ask me the questions you need to be useful to me. One at a time. Then tell me which of your memory files look stale.”
In this setup, memory includes files I can inspect and correct. I ask the agent to show what it used, then check each proposed change against the evidence.
NextAGENTS.md. The file that points it at everything else.
AGENTS.md — my most important instruction file
My shared rules, loaded by each configured tool
I keep shared rules in one canonical file and configure my tools to load it. Claude Code uses CLAUDE.md; Codex uses AGENTS.md. Project instructions add local rules.
NextThe skills it routes to.
Skills: ways of working, written down
Six from my setup
A skill packages instructions and may include scripts or references. The agent reads the selected skill and follows its workflow. I can inspect and change that workflow.
NextOpen the orchestrator skill and run it.
Give the orchestrator a job
“Use orchestrator. From this meeting note, give one worker the actions and another the unresolved decisions. Show me the plan and routes before dispatch.”
I give it distinct jobs. It chooses a route by difficulty and how the result can be checked, then checks provider allowance. Each worker receives enough context to do its part.
NextThe files those agents work from.
Vaults — your knowledge, in files
The five vault kinds I use
Obsidian stores notes as local Markdown files, alongside attachments and settings. An agent with file access can read those notes directly. I organise vaults by what the work needs.
NextThe agents vault, the one nobody thinks to create.
The agents vault
Why it exists
“A fortnight ago I was working on a proposal about a team workshop. Find it and pull it up.”
My standing rule asks each configured agent to leave a session note. Those notes make earlier work easier to find. Their value depends on recording the goal, result and file paths.
NextA vault you build up rather than work through.
A wiki for the method
The project vault holds one piece of work. The wiki holds the method: how the work is done and the standard it is judged against. I keep the method stable and its evidence current.
NextHow it all connects.
How it all connects
The harness puts the model to work. Context engineering shapes what it knows. Graph engineering defines jobs, dependencies and acceptance checks, so independent work can run in parallel.
NextBecome an engineer by Christmas.
Become an engineer by Christmas
Four levels. Start at the first.
Each level adds one thing: the tools, then HTML, then live data from outside, then a system that runs on its own. Choose the level you can finish.
NextWhere to find all of this.
Take it with you
The guide
Every tool on this page, written up
barnabyrobson.org/resources/ai-tools
My skills
Eight skills and a starter AGENTS.md, free to copy
github.com/barnabybot/public-skills
The guide explains the tools and the harness. The repository holds eight of my skills and a starter AGENTS.md, each rewritten to run on its own. Copy one and make it yours.