Vibeleaderboard
← Course overviewModule 04 · Give the agent the right information · Beta
Lesson 4.1

Module contents

Give the agent the right information

Lesson 1 of 5

What does the agent need to know, and where does it live?

Module 3 ended with a delegation brief that was deliberately incomplete. It said nothing about who needs the fencing comparison, by when, or where the Garden folder is. This lesson starts from those gaps. It sorts the information a task needs into four kinds by purpose, shows what a missing fact does to a run that follows the brief, and gives the question the rest of the module keeps asking: is this fact where the harness can use it, at the moment it is needed, and is it right?

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

Four kinds of information, sorted by purpose

A model brings its learned values to every task, and module 1 explained what those can and cannot supply. Anything specific to this task, this project or this organizer has to be supplied in the context or retrieved by a tool. Module 1 introduced the as the bounded space that holds the request, any supplied material and the conversation so far, and module 3 added tool descriptions and observations to it. Sorting that material by purpose is more useful than sorting it by where it came from, because the purpose tells you how long it stays relevant and where it ought to be kept.

A is information relevant to one particular task: which quotes to compare, that the committee needs the table, that the meeting is on Thursday. It does not stop being true when the task ends; it may simply not apply to the next task, and that is what decides where it is kept. Source evidence is what the task is about and what the tools return: the quotes themselves, a folder listing, a fetched page. Module 3 called the returned kind an . A is a rule or fact that applies across the tasks in one project: prices are compared installed, the Garden folder is inside Documents on the organizer's laptop, anything for the committee goes in a folder named Committee. Conventions can change, and they can have exceptions, so a convention is a standing default rather than a law. A is a written way of doing a recurring job, such as the steps for reviewing any supplier quote, which apply again next month for the pond pump.

These kinds differ in how long they stay relevant. A task fact is relevant for its task. A convention is relevant across the project, until the project changes it. A procedure is relevant whenever its job comes up. Evidence is whatever the tools returned this time. Keeping a fact in the wrong place is a failure you can see coming. A deadline written into a file every task reads is still a true statement about one meeting and a misleading cue for tasks it does not apply to. A folder location repeated in each brief will eventually be left out of one.

The brief's three gaps, and what a model does with a gap

The module 3 brief was missing two task facts and one convention. Who needs the comparison and by when are task facts; where the Garden folder lives is a convention. Here is an authored run with the original brief, given to the same fictional harness and set up to match the brief as module 3 described: reads allowed in Garden/Suppliers, one write to Garden/fencing-comparison.txt, no email tool, and a stop condition for anything that would need an action not listed. Module 2 introduced the notation: Garden/Suppliers means the folder Suppliers inside the folder Garden, counted from wherever the harness started, which here is the organizer's home folder on their laptop.

 1  [model: call]    list_folder   Garden/Suppliers
 2  [observation]    error: no such folder: Garden/Suppliers
 3  [model]          The brief names Garden/Suppliers, but no folder exists at that path from here.
                     Finding it would mean reading outside the allowed folder, so I am stopping.
                     Where is the Garden folder?
 4  [person]         It is Documents/Garden. One moment.
 5  [harness]        Rules updated by the person: reads allowed in Documents/Garden/Suppliers;
                     write allowed to Documents/Garden/fencing-comparison.txt.
 6  [person]         Go ahead with the brief, using Documents/Garden as the Garden folder.
 7  [model: call]    list_folder   Documents/Garden/Suppliers
 8  [observation]    fencing-brief.txt   northfence-quote.txt   hedgerow-quote.txt

Line 3 is the brief working as written. The path in the brief did not exist from where the harness started, and finding the folder would have meant listing folders the brief did not allow, so the model stopped and asked instead of searching. Line 4 supplies the missing convention. Line 5 is the part a chat message cannot do: the harness's rules named the old paths, so the person changes the rules before the new read can run, and line 6 restates the task. From line 8 the run continues as module 3 traced it, with the same reads, the same two timeouts and the same stop condition. The cost of the gap was one stop, one answer and one settings change. Had the harness been set to allow reads anywhere, a model might have searched the laptop instead, which finds the folder faster when it finds the right one and can go wrong in several ways when there are two folders named Garden.

The other two gaps produce no error line at all. Nothing in the run mentions the committee or Thursday, because nothing told the model they exist, and whatever is written will be saved in Garden rather than in the committee's folder. Module 1 described what a model does with a gap: it generates from what it has. Sometimes it asks, as on line 3. Sometimes it can retrieve the fact with a tool, if the fact is somewhere a tool can reach. Sometimes it fills the gap with a plausible guess, which module 1 called a when the guess is presented as fact. A more capable model may ask better questions and search more sensibly. What no model can do is know a private requirement for which it has no evidence: that this particular committee meets on Thursday exists nowhere the model can look. Writing the brief reduces the gaps you have identified. It does not rule out gaps you have not.

Each gap belongs somewhere different. The committee and Thursday belong in the brief, because they are relevant to this task. The folder location belongs with the project, in a place every garden task reads, because it is relevant to all of them. Lesson 3 shows what that place is in real harnesses.

Choosing what goes in

Copying the whole garden archive creates a different problem: the quotes arrive alongside twenty years of unrelated minutes. Anthropic’s 2025 context-engineering article recommends selecting relevant information as work proceeds. 8 The next lesson explains why fitting information into a window does not guarantee using it well.

Relevance is one test, and it is not the only one. A piece of context also has to be accurate, current and from a source you would accept: a folder location that was right last year, a price basis the committee has since changed, or a quote copied from the wrong supplier is relevant and wrong, and the model may have no evidence that reveals the error. So the question to ask of each piece is: does the model need it for this step, is it present at that step, and is it right?

The answer takes a different shape for each of the four kinds. Task facts go in the brief, and only the ones this task needs. Evidence can be supplied up front or fetched by a tool when a step needs it; which is better depends on how much there is and whether it is all needed. Conventions are usually kept somewhere the harness reads at the start of every task in the project, so that no brief has to repeat them. Procedures are best kept where they can be brought in for the job they describe and left out otherwise. These are arrangements harnesses make possible, not rules, and the rest of the module is about the mechanisms behind each of them and about what gets lost between them.

Instructions are information too

Module 3 placed instructions in the first of its three layers: an instruction shapes what the model proposes, and a harness rule or an enforced boundary decides what runs. This lesson adds the other half of that fact. An instruction is a piece of context, subject to everything above. It has to be present to have any effect at all. It draws on the same limited attention as the quotes and the observations. It can be out of date, if it was written for a task, a project rule or a model that has since changed. And it can conflict with another instruction that is also present, in which case the model has to resolve the conflict somehow, and you may not like how.

So the four-way sort applies to instructions as well as to facts. "Compare installed prices" is a convention. "Stop and report if a quote cannot be retrieved after one retry" was a task instruction in the brief. "For each quote, check the length, the price basis, the delivery time and the guarantee, in that order" is a procedure. When someone says an agent has instructions, ask the same questions as for a fact: where is it kept, when is it loaded, and is it still right.

Carry this forward

  • Sort what a task needs by purpose: task facts, source evidence, project conventions and reusable procedures. Each stays relevant for a different span, which decides where it should be kept. A task fact does not become false when its task ends; it may simply not apply to the next one.
  • A fact the model never received cannot shape what it proposes. The model may ask, retrieve or guess; a more capable model may ask and search better, but no model can know a private requirement it has no evidence for. A brief reduces the gaps you identified, not every possible gap.
  • Ask of each piece of context whether the model needs it for this step, whether it is present then, and whether it is accurate and current. Relevance and truth are separate tests, and more context is not automatically better context.
  • An instruction is context: it must be present, it draws on limited attention, it can be out of date, and it can conflict with another instruction.

Lesson check

The organizer writes "The Garden folder is Documents/Garden" into the brief for the fencing task and, next month, forgets to write it into the brief for the pond pump task. Which sort of information is that sentence, and what does the lesson say about where it belongs?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

The four kinds now have names. The next lesson asks why a fact that is present can still be missed, and what happens when two instructions in the same context disagree.

Sources · Reviewed 2026-09-29
  1. 8. Anthropic: Effective context engineering for AI agents (September 2025)

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.