Lesson 2.1 — The context window is a desk

It's a desk, not a filing cabinet — only so much fits, and new papers bury old ones.

TL;DR: The context window is finite working space the agent actively reasons over — not storage it can query. Everything competes for the same token budget, and most of what lands there is incidental.

ELI5: desk, not filing cabinet

The agent works at a small desk, not an endless cabinet.

	Filing cabinet	Desk (the context window)
Size	Effectively infinite	Finite — a fixed token budget
Access	Query any file later	Only what's physically on it right now
When full	Add another drawer	New papers bury old ones; nothing's found

The context window is the desk. Everything the agent "knows" right now has to fit on that desk. It is not a database it can query — it's the finite surface it's actively looking at ¹.

   ┌───────────────── THE DESK (context window) ─────────────────┐
   │ system prompt | AGENTS.md | files you opened | tool outputs │
   │ | every message in this conversation | the agent's replies  │
   └─────────────────────────────────────────────────────────────┘
        ↑ all of this competes for the same finite space

🧠 Test Yourself: If the agent read a file 20 turns ago and "forgot" a detail in it, where did that detail go?

Answer

Nowhere it can retrieve on its own — it's still on the desk, just buried/diluted among everything since. The window isn't searchable storage ¹.

What's actually on the desk

Five things eat the same budget; files and tool output are the silent hogs.

#	What's on the desk	Notes
1	System prompt	The agent's built-in instructions — always there.
2	Steering files	AGENTS.md / CLAUDE.md / .cursor/rules — loaded each session.
3	Files the agent read	Full contents — even the parts it didn't need. ⚠️ silent hog
4	Tool output	Test logs, grep results, build errors, stdout. ⚠️ silent hog
5	The whole conversation	Your messages and every prior agent reply.

Rows 3–4 are the budget-eaters. Ask the agent to "read the codebase" and it can drop tens of thousands of tokens of file contents onto the desk — most of it irrelevant. Anthropic notes tool results alone can consume 50,000+ tokens before the agent even starts on your request ¹.

Worked example: watch the desk fill

One bug fix, five turns, desk goes from 5% to buried.

flowchart TD
    S["Start: system prompt + AGENTS.md — ~5% full"]
    T1["Turn 1: greps 12 files, reads 6 fully — ~35% full"]
    T2["Turn 2: runs test suite, long failing output — ~50% full"]
    T3["Turn 3: you paste a stack trace, reads 3 more files — ~70% full"]
    T4["Turn 4: two failed fix attempts, dead-end code — ~85% full"]
    T5["Turn 5: 'why is it getting confused?' — the desk is buried"]
    S --> T1 --> T2 --> T3 --> T4 --> T5
    T5:::buried
    classDef buried fill:#fee,stroke:#c00,stroke-width:2px;

Nothing went wrong with the model. The desk filled with junk — six fully-read files it no longer needs, a giant test log, and two abandoned approaches still in view.

🧠 Test Yourself: By Turn 5, what's the single biggest category of waste on the desk?

Answer

The fully-read files and the long test log (rows 3–4) — incidental content the current task no longer needs. That's what you'll learn to clear in Lesson 3.

Why this matters

You can't manage what you can't see. Stay aware of how full the desk is, and with what.

Before the moves in Lesson 3 make sense, internalize this: the agent always reasons over a finite, fillable surface, and most of what lands there is incidental. Good operators keep a constant background awareness of "how full is the desk, and with what?" ¹.

Your turn (exercise)

In your next session, after ~10 turns, ask the agent:

"Roughly what's taking up most of your context right now — files, tool output, or our conversation?"

Then identify one thing on the desk the current task doesn't need. That noticing is the skill. (Lesson 3 is what you do about it.)

---

← Phase 2 home · next → Lesson 2.2 — Context rot

---

1. Effective context engineering for AI agents — Anthropic ↩↩↩↩