Four words. Everyone at THG uses them. Almost nobody uses them correctly. This module fixes that in ten minutes, and then makes you prove it.
~10 min1 drag game1 quickfirequiz · pass ≥ 80%
THE BRUTAL TRUTHIf you can't tell a model from an agent, you will blame the wrong thing when it breaks. Then you will keep breaking it, and you'll call it "the AI being stupid". It isn't. You are.
The four things, defined so you can't get them wrong
Read each one once. There is no "it depends". These are the definitions we use at THG, in prompts, in tickets, in Slack. Use them.
Model
The brain in a box. Text in, text out. Nothing else.
Opus 5, GPT‑5.6, Sonnet 5 are models. A model has no memory between calls, no hands, no internet, no idea what happened yesterday. It cannot run anything, open anything, or delete anything. It predicts the next words based on what is in front of it right now.
AT THG THIS ISYou paste an RTP question into a chat window and get an answer. That's a model talking. If it made up a GLI clause, that's a model doing what models do when the source isn't in front of them.
Agent
A model plus a loop plus tools plus a goal. It acts.
An agent reads, decides, does something real (runs a command, edits a file, calls Jira, sends a message), reads the result, and goes again until the goal is met or it gets stuck. It is the only one of the four that can actually break production.
AT THG THIS ISClaude Code editing the Unibit front‑end. Tico orchestrating the daily backup. An OpenClaw automation ingesting Zoom recordings. Give an agent write access and a vague goal and you have given a stranger your keys.
Subagent
A worker the agent hired for one job, with a clean desk.
An agent spawns a subagent to do a scoped task: review this PR, research this spec, test this build. The subagent starts with fresh, empty context. When it finishes, it reports back a summary and everything else it knew is gone. That's the point: fresh eyes, no contamination from the builder's assumptions.
AT THG THIS ISMain agent builds the crash‑game payout table, then spawns a subagent to review it. The reviewer never saw the builder's reasoning, so it catches what the builder rationalised away.
Context
The desk. Everything the model can see right now. It fills up.
Your prompt, the files loaded, the conversation so far, every tool output: that is context. It is finite. Long before it is "full", quality drops. Old details get compressed or fall off. Context is not memory. Close the session and the desk is wiped clean.
AT THG THIS ISYou paste 40 pages of GLI‑19, ask fifteen questions in a row, and wonder why answer fifteen contradicts answer two. The desk is buried. Nobody's fault but the person stacking paper on it.
Model = thinks. Agent = acts. Subagent = fresh pair of eyes. Context = the desk that fills up. Say it out loud once. You now know more than most of LinkedIn.
Spot the difference: who actually failed?
Every "the AI screwed up" complaint at THG maps to exactly one of these. Get the diagnosis right and the fix is obvious. Get it wrong and you'll "fix" it by yelling at the chat window.
"It invented a GLI test case number that doesn't exist."
model
Models fill gaps with plausible text. Put the real spec in context or give it a tool to look it up. Never ask a bare model for a fact you can't verify.
"It deleted the wrong branch."
agent (and you)
Only an agent can delete. You gave it the tool and a goal loose enough to include deleting. Scope the goal, take away the tool, or make it ask first.
"Two hours in, it forgot we agreed to keep the old API."
context
The decision scrolled off the desk. Re‑state the plan. Better: one small task per session, plan written down, referenced every prompt.
"The review missed the exact bug the builder introduced."
missing subagent
You reviewed in the same session that built. Same desk, same blind spots. Spawn a subagent with fresh context to review. Always.
"It keeps asking me obvious questions."
you
It doesn't have what you think it has. Nothing is on the desk unless you put it there. Under‑specifying is your error, not the model's.
Don't say this"The AI is broken again, it forgot everything." Nobody can act on this. Which "AI"? Forgot what? Where?
Say this"The agent's context is saturated after 3 hours of building. I'm starting a fresh session with the plan pasted in, and spawning a subagent to review what's done." Diagnosis, cause, fix. Ondrej reads this and nods.
Game 1: drag the term onto its definition
Six terms, six definitions. Drag a term onto the matching definition. Wrong goes red and shakes. Right goes green and locks. On touch or keyboard: tap or press Enter on a term to pick it up, then tap or press Enter on a definition to drop it.
Try to get 6 out of 6 with zero misses. Misses are counted. Real life counts them too, in hours.
Match the terms
Placed 0/6Misses 0
Game 2: quickfire, which one is it?
Real THG situations. Which of the four is the actual subject? You get one shot each. Build a streak.
Which one is it?
Score 0/0streak 0
Situation 1
Watch it happen: one session, three failures, three different causes
Press play. This is a real shape of a real bad afternoon. Note which word explains each red line.
thg‑dev · agent session · 14:02 → 17:40
press ▶ to run the afternoon
Quiz: prove it
Five questions. You need 4 right (80%) to pass. Fail and you retake it. Nobody is watching, but the progress record is.
Module 1 quiz
THE ONE THING TO REMEMBERThe model thinks. The agent acts. The subagent is a fresh pair of eyes. The context is a desk that fills up. Before you blame "the AI", name which one failed. If you can't, the failure is yours.