0x20Lesson 3 of 8

Track state, progress and context

Checkpoint progress so work can resume, and keep the context window focused with compaction, notes and just-in-time retrieval.

22 min 6-question quiz 2 code exercises

By the end of this lesson you can

Design explicit, durable progress state
Resume after an interruption without repeating work
Manage a long run’s context with compaction and notes

A multi-step loop needs state: the goal, constraints, completed steps, key results and what’s next. Keep it explicit, compact and durable - saved outside the model’s context - so a crash, timeout or restart doesn’t lose an hour of work. The model can propose a state update; your code validates it before saving.

Long runs also fight context rot: as the context fills up, models recall details less reliably (Anthropic, “Effective context engineering for AI agents”). Techniques:

Compaction - summarize old history and drop stale tool output, keeping decisions and recent details.
Structured notes - the agent keeps a progress file or to-do list outside the context and rereads it.
Just-in-time retrieval - keep references (paths, IDs, queries) and load content only when needed.
Sub-agents - delegate focused work to agents with clean contexts that return short summaries.

Try it

Up to 15 XP

What stays in the context?

A long coding session. Switch window sizes: what gets dropped? Which parts would you compact into a summary, and which would you keep as notes outside the context?

Context window:

System prompt + tool docs

Design decisions so far

Progress notes (to-do list)

Last 5 steps

Current instruction

answer

Using 23,800 of 32,000 tokens (4,000 reserved for the answer) · the full conversation is 94,800 tokens.

System prompt + tool docs6,000
Early exploration (file listings)30,000
Test output from 40 runs45,000
Design decisions so far3,000
Progress notes (to-do list)1,500
Last 5 steps9,000
Current instruction300

Dropped to fit: Early exploration (file listings), Test output from 40 runs. The model will never see them - it can’t answer from what was cut.

checkpoint.py

1import json
2state = {"goal": "migrate 3 tables", "done": ["users"], "next": "orders"}
3saved = json.dumps(state)
4print(json.loads(saved)["next"])

Output

orders

Key takeaways

Keep explicit, validated, durable state: goal, done, next.
Checkpoint so interrupted work resumes instead of restarting.
Fight context rot with compaction, notes, just-in-time retrieval and sub-agents.

Lesson quiz

6 questions · pass with 5 correct · up to 50 XP

Passing this quiz completes the lesson and keeps your streak going. Questions you miss come back in review sessions later.

Practice: write Python

Write Python in the editor and run it against sample inputs. Python runs locally in your browser using a WebAssembly runtime.

Exercise 1

Resume from a checkpoint

+25 XP

Line 1 is a JSON checkpoint {"done": [...]}. Line 2 is the JSON plan (a list of step names). Print resuming: D/T steps already done, then skip STEP or run STEP for each planned step, then done: T/T.

Two of four done

main.py

Loading editor…

Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.

Exercise 2

Compact a long history

+25 XP

Line 1 is N (at least 1). Each following line is a message role: content. Keep the system message(s) and the last N other messages verbatim. Replace everything older with one line summary: K earlier messages compacted (T tool results dropped), placed right after the system messages. Print the result (omit the summary line if nothing was compacted).

Keep the last two

main.py

Loading editor…

Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.

Questions about this lesson

Stuck? Ask. Figured something out? Share it. Explaining is one of the best ways to learn.

Loading posts…