# When to stop delegating and do it yourself

> Read three round logs, name the three signs that another round costs more than doing the step yourself, and see why taking a step back says nothing about the tool or about you.

Page: https://schubergphilis.github.io/ai-training/using-agents/when-to-stop/
License: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/

In this lesson we read three logs of delegated work and find, in each, the
round where handing the piece back to the agent started to cost more than
doing the step yourself. Then we name the signs that show that point, and
we look at why taking a step back is a judgment about cost and says nothing
about the tool or about you. You don't run anything in this lesson. The
logs are supplied, and the work is reading them.

A round, in this lesson, is one pass of delegated work. You hand the agent
a piece with a brief and get a result back. After checking the result
against the done-criteria, you change the brief with what the check showed
and hand the piece over again. Working in rounds is how a split task gets
done, and most pieces take one or two. This lesson is about the piece that
doesn't.

## The round logs

The logs below come from the calibration team at Brindlewick Instruments,
the fictional company from the earlier lessons. Each log is a few lines per
round: what was handed over, what came back, what the check found. Read them
once before the next section names what they show, and keep in mind that
each one stops for its own reason.

The first log is the "Blocked" section of the weekly status update. The
brief for this piece said: list every item from the notes that's waiting on
someone outside the team, one bullet each, with who it is waiting on and
the date.

```text
Log A: the Blocked section

Round 1
  Handed over: the brief and notes.md.
  Came back: two bullets. The reference weight, waiting on the supplier.
    The thermocouple batch, waiting on the food lab.
  Check: the thermocouple batch is in progress, and the team is doing it.
    The notes say "12 of 40 done", and nothing is waiting on the lab.
  Brief change: added "an item counts as blocked only when the team
    cannot continue until someone outside the team acts".

Round 2
  Handed over: the changed brief.
  Came back: one bullet, the reference weight. A second bullet: the
    spirit levels, waiting on the survey team to confirm the start date.
  Check: the notes say "start the spirit levels" under next week. There
    is no confirmation in the notes. The agent read a plan as a wait.
  Brief change: added "use only what the notes say is waiting".

Round 3
  Handed over: the changed brief.
  Came back: one bullet, the reference weight, "due 26 Sep, supplier".
    A note under it: "the storeroom probe shortage may also block
    calibration work".
  Check: the bullet is right. The note is a guess, and the brief asked
    for bullets only.
```

The section is one bullet long. The person running this log has now read
three results, found a wrong line in each, and written two additions to the
brief. Writing the bullet by hand takes less time than reading round three.

The second log is a piece of the same update that the director asked for
after the meeting: mark which of the finished items she has already signed
off, so the update doesn't ask her to look at them again.

```text
Log B: the signed-off marks

Round 1
  Handed over: the brief, notes.md and the finished update.
  Came back: the update with "(signed off)" on none of the items, and a
    question: "The notes and the update do not say which items the
    director has signed off. Where can I find that?"
  Check: the question is right. The sign-offs happened in a meeting,
    and nothing in the folder records them.
  Brief change: added "the torque wrenches were signed off in the
    meeting of 18 Sep".

Round 2
  Handed over: the changed brief.
  Came back: "(signed off)" on the torque wrenches. A question: "Was the
    storeroom count also signed off? It is listed as done."
  Check: it was, and again the answer is in your memory of the meeting.
```

Every round of this log ends with the agent asking you for a fact and you
typing it. The agent adds nothing to the fact. The only place the answer
exists is your head, so the step is yours whether you delegate it or not.

The third log is the pressure gauge batch the team calibrated last month.
The team's own calibration sheet lists 40 serial numbers with a pass or
fail for each, and the piece is a two-line count for the update: how many
passed, how many failed.

```text
Log C: the pass and fail count

Round 1
  Handed over: the brief and the calibration sheet, pasted in.
  Came back: "37 passed, 3 failed."
  Check: the sheet has 40 rows. Counting the fails by hand gives 4,
    and one of the rows says "fail (recheck)" in a different column.
  Brief change: added "a row with fail anywhere in it counts as a fail".

Round 2
  Handed over: the changed brief and the sheet.
  Came back: "36 passed, 4 failed."
  Check: read all 40 rows again, one by one, against the two numbers.
```

The result is right in round two. Getting to "right" took reading 40 rows
twice, and reading them once with a tally is the whole task. When the check
is the work, delegating it doesn't remove the work from your desk. It
doubles it.

## The signs

Each log ends in a different round, and each shows one sign that it was
time to take the piece back.

**A third round on one piece.** The first result often shows a gap in
the brief, and filling the gap for a second round is the normal
course of working in rounds. By the third round the second fix hasn't held
either, and the piece has become a series of small corrections on
something you could see the answer to from round one. In Log A that's the
third result on a one-bullet section.

**A step that depends on knowledge only you hold.** The agent works from
what it can read. When the fact the piece needs is in your memory, in a
conversation you had, or in a judgment only you are in a position to make,
every round is the agent asking and you answering. In Log B the sign-offs
happened in a meeting and are written down nowhere the agent can reach.

**A check that takes longer than the work.** A result you can only trust
after redoing the task is a result you have done yourself, with a
delegation on top. In Log C the check is reading every row, and reading
every row once is the piece.

The signs are silent about how well the agent did. The agent in Log A wrote a
correct bullet in every round and a wrong extra each time. The agent in Log
B asked the right question twice. The agent in Log C got the count right
once the brief said what a fail looks like. The signs measure the cost of
the next round against the cost of the step. That cost is a fact about the
piece and the situation, and it can be read from the log.

#### Checkpoint: Which sign does each log show?

Match each log to the sign it shows.

Statements:

- Log A, the Blocked section
- Log B, the signed-off marks
- Log C, the pass and fail count

Options:

- A third round on one piece
- A step that depends on knowledge only you hold
- A check that takes longer than the work

## Taking the step back is a cost judgment

When a sign shows, take the piece back and do the step. The reason is
cost. A round has a price: writing the brief change, waiting for the
result, and reading it against the done-criteria. A step done by
hand has a price too. When the next round costs more than the step, the
step is cheaper by hand, and that is the whole argument. Which tool you
used, and how good it is, doesn't enter into it. Neither does how good you
are at briefing. A brief can be written well and still be handed to a
piece where the fact needed is in your memory, and no rewrite of the brief
changes where that fact is.

Taking one piece back also loses nothing else. The piece is still the piece
your split produced, with its own input, its own output and its own check.
The pieces around it stay delegated, and the check you wrote for this piece
still applies to the result you make by hand. In Log A the bullet still has
to name who it waits on and the date. In Log C the two numbers still have to
add up to 40. The split shaped the work, and the split holds whether an
agent or you fills a given slot in it.

The judgment runs the other way as well. A piece that shows none of the
signs stays with the agent even when doing it yourself would be quicker
today, because the brief you write is one you can run again next week, and
a step you did by hand is gone when it's done. Delegating is a judgment
about what the situation makes cheap, and the same person delegates one
piece and takes back the next (AI Fluency: Framework and foundations, Claude Academy).

#### Pitfall: Finishing the wrong thing well

A learner reaches a third round on a one-bullet section and feels that
stopping now would waste the earlier rounds, so they change the brief once
more and run a fourth. The fourth result is clean, and four results and
three brief changes went into one bullet. The earlier rounds cost what they
cost, and a fourth round adds to that cost without taking any of it back.
The feeling that stopping is giving up is about you, and the log is about
the piece, so read the log and take the step yourself when it shows a
sign.

#### Checkpoint: The third result is back

The third round on a short list comes back with one right line and one
wrong extra line, as the two before it did. The list is four lines long.
What do you do?

- Write the list yourself, check it against the done-criteria in the brief, and move on to the next piece.
- Add one more line to the brief that rules out this extra, and run a fourth round, since three rounds of work would be wasted otherwise.
- Switch to a different agent and hand it the same brief, since this one keeps adding a line.
- Accept the result with the extra line and mark the piece done, since it has a right line in it.

## Exercise

Below are three short round logs from the same fictional team. For each
one, write down the round where taking the step back would have been
cheaper, and the sign that shows it. Then write one line on what the person
would have done by hand at that point. Ten minutes is enough.

```text
Log 1: the meeting date line

Round 1
  Handed over: write one line naming the next review meeting, from
    calendar.md.
  Came back: "Next review: Thursday 2 Oct, 10:00."
  Check: calendar.md has the meeting on 2 Oct. The time moved to 14:00
    in a message to you this morning, and the file still says 10:00.
  Brief change: added "the meeting is at 14:00".

Round 2
  Handed over: the changed brief and calendar.md.
  Came back: "Next review: Thursday 2 Oct, 14:00."
  Check: correct. A question: "Should I also note that the room
    changed? calendar.md lists room 3 and the message you mention
    may say otherwise."
```

```text
Log 2: the supplier list

Round 1
  Handed over: from the six order emails pasted in, list each supplier
    once with the item ordered.
  Came back: six suppliers, six items.
  Check: two of the emails are from the same supplier. The list has it
    twice, under two spellings of its name.
  Brief change: added "the same supplier under two spellings is one
    supplier".

Round 2
  Handed over: the changed brief and the six emails.
  Came back: five suppliers, six items.
  Check: correct. A seventh line lists an item with "supplier not
    named". The email for it names the supplier in the signature.
  Brief change: added "the supplier may be named in the signature".

Round 3
  Handed over: the changed brief and the six emails.
  Came back: five suppliers, six items, and a note that one email may
    be a duplicate of another.
  Check: the emails are not duplicates. The list is right apart from
    the note.
```

```text
Log 3: the serial number check

Round 1
  Handed over: the sheet of 40 serial numbers from the calibration
    bench and the sheet of 40 from the storeroom. Report any number on
    one sheet and not the other.
  Came back: "All 40 numbers appear on both sheets."
  Check: read both sheets side by side, 40 rows each, and found one
    number on the bench sheet with two digits swapped.
  Brief change: added "compare digit by digit".

Round 2
  Handed over: the changed brief and both sheets.
  Came back: "One mismatch: row 23, digits 5 and 6 swapped."
  Check: read both sheets side by side again to confirm there is no
    second one.
```

A good answer names Log 1 round 1, because the check shows that the fact
the piece needs arrived in a message to you and is in no file the agent
can read, and the question in round 2 shows the same thing again. It names
Log 2 round 3, because the piece is a five-line list and the third
result again holds one thing the brief never asked for. It names Log 3
round 1, because trusting either result means reading both sheets side
by side, and that reading is the task. Where your answer differs, write
down which cost you weighed differently. Which piece in your own work
this week reached one of these three points?

Stretch: Take a piece from your own recent work that you delegated and re-ran more than once. Write its round log in the format above, and mark the round where you would stop now.

## Recap

1. A round is a brief handed over, a result checked against the done-criteria, and a brief changed with what the check showed. Most pieces take one or two, and a first result with a gap is the normal shape of a round.
2. A third round on one piece, a step that depends on knowledge only you hold, and a check that takes longer than the work are the signs that the next round costs more than the step done by hand.
3. Taking the piece back is a judgment about cost (AI Fluency: Framework and foundations, Claude Academy). It says nothing about the tool, its quality, or your briefing, because none of those move where a fact is or how long a check takes.
4. The piece you take back keeps the input, output and check the split gave it, and the other pieces stay delegated, so the split loses nothing.
5. Stopping after spent rounds can feel like giving up. The spent rounds cost the same either way, and the log says whether the next step is yours, however stopping feels.

## References

- AI Fluency: Framework and foundations, Claude Academy, https://academy.claude.com/courses/ai-fluency-framework-foundations