Part 5 · When the Harness Improves Itself

Three Endurance Rules

A context-usage curve and compression-threshold demo; three tasks: find the forgetting turn, write three rules, run a write-to-disk-then-read-back loop

THE QUESTION THIS PAGE ANSWERS

ANSWER FIRST

What is the key idea behind “Three Endurance Rules”?

A context-usage curve and compression-threshold demo; three tasks: find the forgetting turn, write three rules, run a write-to-disk-then-read-back loop

DECISION RULE

Follow the handoffs, not the demo. A system becomes dependable at the boundaries between model, tools, state, permissions, and people. Read each handoff as a place where you can observe, test, and recover.

TRY NEXT

Name the input, owner, approval, and recovery action for one automated step.

WATCH FOR

A successful run that cannot explain what happened or be safely repeated.

Fifth square on the practice track: you have sprint results. Now train for the marathon.
M0
Know what you want it to do for you
M1
Talks like a human, reliably
M2
Actually gets work done
M3
Better or worse, measurably
M4
Runs long without losing memory
M5
The process is reproducible
See it first · Forgetting is a curve you can draw

When a user says "it got dumber as we talked," what's behind that is a very concrete curve: context usage climbs turn by turn, and once it's full the system can only quietly throw things away. What it throws is usually the earliest messages — the most important rules you laid down at the start.

20 turns of continuous chat: how context usage moves

Context usage 85% compression threshold
Do nothing to manage it, and it hits the threshold on turn 14. Every turn after that quietly drops the earliest messages. What the user feels is "it forgot what I said at the start."

Three rules, three kinds of forgetting

Compression threshold handles "only act when it blows up": compress when usage hits the line. Don't wait for an error.

Never-delete list handles "compressed too hard": the user's original words, the agreed requirements, the safety constraints — once those go into a summary you can't get them back. They must stay pinned in context.

Notes on disk handles "forgetting across sessions": write important conclusions to a file, and read the file first next time you start. That's what this chapter calls "filesystem persistent memory," and it's the starting point of Harness self-improvement. It has to remember what it learned last time before it can improve itself.

Hands-on checklist · Pick one, start it, tick it off

This Chapter's Hands-on Checklist

0 / 3 done

Find the turn it starts forgetting

15 min Everyone

Pull up your longest conversation with an AI. Read it from start to finish. Find the first turn it forgot something you had already told it. Write that number down — that's your Agent's real endurance with no management at all.

What counts as done
You can say "around turn N, it forgot XX." If you can't, your conversations aren't long enough yet. That's a useful conclusion too: your scenario doesn't need compression for now. Don't over-engineer it.

Write your three endurance rules

1 hour Anyone whose Agent runs long tasks

Make the three rules concrete for your scenario: compress when usage hits what percent, which content never enters the compression zone (make a list — don't write vague junk like "important content"), and which conclusions get written to a file. Then paste them into your Agent config or system prompt.

What counts as done
All three have a concrete number or list, and the "never-delete list" at least includes: key constraints from the user's own words, and the requirement definition set in M0.

Run one "write to disk, then read it back"

Half a day Anyone who works across sessions

After the Agent finishes a piece of work, have it write the conclusions into a fixed file (call it a work note). Then close the session, open a brand-new one, and as the first message tell it to read that file and continue. If the new session picks up seamlessly, you've just implemented this chapter's structured-notes pattern — the simplest and most useful form of long-term memory.

What counts as done
In the new session it doesn't ask any "where did we leave off last time" fluff — it continues from the breakpoint. Once you can do this, M4 is standing.

Save the three rules in the Build Log

What M4 needs to record: the forgetting turn, the three rules in full, and where the on-disk notes live. The next square is the last one: turn the whole process into a spec someone else can reproduce.

Fill in M4

How “See it first · Forgetting is a curve you can draw” becomes executable

“When a user says "it got dumber as we talked," what's behind that is a very concrete curve: context usage climbs turn by turn, and once it's full the system can only quietly throw…” is not about a magic phrase. It is about giving the model enough information to know who the work is for, what must be done, and what counts as acceptable.

Background sets direction; constraints set the boundary

“Compression threshold handles "only act when it blows up": compress when usage hits the line.” shows why a useful request separates the task, audience, source material, output format, and constraints. Without background, the model guesses. Without acceptance criteria, fluent text is not evidence that the task is complete.

More words do not guarantee a better result

Turn “What M4 needs to record: the forgetting turn, the three rules in full, and where the on-disk notes live.” into a small experiment: change only one of background, requirements, or constraints while keeping the rest fixed, then observe which layer actually changes the output.

From “See it first · Forgetting is a curve you can draw” to “Hands-on checklist · Pick one, start it, tick it off”

“See it first · Forgetting is a curve you can draw” grounds the problem in “When a user says "it got dumber as we talked," what's behind that is a very concrete curve: context usage climbs turn by turn, and once it's full the system can only quietly throw things away . What it throws i…”. “Hands-on checklist · Pick one, start it, tick it off” then moves it toward “Pull up your longest conversation with an AI. Read it from start to finish. Find the first turn it forgot something you had already told it . Write that number down — that's your Agent's real endurance with no…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.

Carry the judgment into the next situation

Build a request layer by layer: task and audience first, material and output rules next, constraints and acceptance checks last. Change one layer at a time so you know what actually helped.

  • “See it first · Forgetting is a curve you can draw”: When a user says "it got dumber as we talked," what's behind that is a very concrete curve: context usage climbs turn by turn, and once it's full the system can only quietly throw things away . What it throws i…
  • “Hands-on checklist · Pick one, start it, tick it off”: Pull up your longest conversation with an AI. Read it from start to finish. Find the first turn it forgot something you had already told it . Write that number down — that's your Agent's real endurance with no…
  • “The closing point”: What M4 needs to record: the forgetting turn, the three rules in full, and where the on-disk notes live. The next square is the last one: turn the whole process into a spec someone else can reproduce

The final “The closing point” brings the discussion to “What M4 needs to record: the forgetting turn, the three rules in full, and where the on-disk notes live. The next square is the last one: turn the whole process into a spec someone else can reproduce”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.

Mark as learned Your reading progress updates automatically
← PreviousNext →

Keep reading

The next useful article in the thread.

ARTICLE DISCUSSION

Leave one useful thought here.

Keep the idea that clicked, the question that stayed open, or a small note for the next learner.

Discussing Three Endurance Rules When the Harness Improves Itself
3discussionsArticle discussion · synced with the Circle
View in the learning circle
AM
Asha MorganContent editor
INSIGHTField note

I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.

ARTICLE DISCUSSION7 helpful
LH
Lin HarperIndie developer
INSIGHTInsight

After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.

ARTICLE DISCUSSION5 helpful
KM
Kiki MooreProduct operations
QUESTIONQuestion

When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.

ARTICLE DISCUSSION4 helpful