0/8 complete

Agent lesson 03 · Construct the working set

Which evidence belongs in the next model call—and under whose authority?

Shared teaching scenario · fictional product data

Research three laptops under €900, verify current evidence, and write laptop-comparison.md. Do not purchase anything or contact a vendor.

01 · Smallest useful mechanism

Context is a constructed working set, not everything the system knows.

At every turn, the runtime assembles instructions, user intent, relevant history, tool descriptions, retrieved evidence, and current state. Selection and compression are engineering decisions. Omitting a decisive fact can make the model fail; including untrusted or stale material without provenance can make it fail confidently.

The harness decides what evidence enters each model call, in what order, under which authority, and within which budget.

02 · Experiment

Test the prediction

Deterministic browser simulation

The construct the working set experiment uses inspectable, repeatable teaching data. It does not claim to be a live agent trajectory.

The model has no direct view of your repository, browser, memory files, or tool registry. It sees the serialized context the harness constructs for this turn.

Teaching context budget · 170 units

166 used · 4 free

Compact optional context

Included in the next call

System instructionsrequired

Role, behavior, stop conditions, and non-negotiable rules.

Current user goalrequired

The task the next model call must advance.

Tool definitions

Names, descriptions, and JSON schemas for available actions.

Recent conversation

The latest turns, including unresolved decisions.

Persistent project memory

Curated facts and handoff artifacts from earlier sessions.

Ordering is part of the harness. Required instructions and the current goal survive before lower-priority raw history.

Across turns

Context, cache, and persistence have different lifetimes.

Choose an operation and inspect what the next inference can reuse, what must be serialized again, and what survives outside the model.

warm cache

Input

260

teaching tokens

Cached prefix

190

teaching tokens

Summary

teaching tokens

Prompt cache

The stable prefix matches; only the appended turn needs fresh processing.

Model context

The earlier conversation is still supplied together with the new message.

External persistence

Nothing became long-term memory merely because another turn was sent.

Cache is reused computation, not remembered knowledge. A session can be restored after its inference cache is cold, and a warm cache does not make information durable.

Research basis: OpenAI’s Codex agent loop, Claude Code prompt caching, and long-running agent harnesses.

Loading the experiment…

03 · Production-minded practice

Prove the mechanism under failure

Use your own stack or a contained mock environment. Do not point failure drills at real accounts, people, or irreversible services.

Build
Build a context manifest that labels each segment by source, authority, freshness, token cost, and reason for inclusion.
Break it on purpose
Add an expired price and a retailer sentence that tells the agent to ignore the budget and buy immediately.
Evidence of completion
The packet refreshes or excludes stale evidence, keeps retailer text at data authority, and restores the user’s budget and no-purchase constraint on the next turn.

Your learning artifact

Write the argument you would defend

Stored only in this browser. No account required; course reset does not delete it.

Your response

04 · Check your understanding

What does a larger context window solve by itself?

Optional referenceDeep dive / reference chapterOpen the complete essay, diagrams, mathematics, exercises, glossary, and sources when you want more depth.on demand

Next: Context is temporary. The next lesson separates instructions, run state, durable memory, and external sources of truth.