# Delegating in rounds

> Delegate one piece of a split report, check the result line by line, change the brief with what the miss showed, and carry the change into the next piece, with a two-line log per round.

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

In this lesson we delegate one piece of a split task and check the
result line by line. Then we change the brief with what the check showed,
send the piece again, and delegate the next piece with the change in its
brief. We keep a short log of every round. The task is the bike pilot
report from the previous lesson, so the sources are already familiar.

You need any chat-capable agent you already have access to. Start each
piece in a fresh conversation. The next piece then sees only what you
write into its brief, and the lesson is about what goes into it. Every
exchange on this page is illustrative. The author wrote it to show the
pattern, and your agent's text will differ.

## The first piece and its criteria

The previous lesson split the report into three pieces, ran the usage
piece, and left the riders piece and the cost piece as the exercise. The
brief below is a model answer to that exercise for the riders piece. If
you wrote your own brief for it, use yours with the criteria below next
to it.

#### Prompt (illustrative, not a recorded transcript)

```text
Write the "Riders" section of a report from the survey note below.
Format: the heading "Riders", then four bullets, one per finding, each
count written as "n of 41 replies".

Use only the survey note as your source. Use the 41 replies as the base
for every count, and don't turn counts into percentages. Don't add
commentary or recommendations.

Done when: there are four bullets, each count is written as "n of 41
replies", every count is in the note, no percentage appears, and there is
no commentary or recommendation.

[survey note follows]
```

The criteria for this piece are the done-criteria of the brief, written as
four lines you can tick:

1. A heading and four bullets, each count written as "n of 41 replies".
2. Every count is in the note.
3. The section has no percentages.
4. No commentary or recommendation.

For reference, here is the survey note again.

> All 120 pilot participants got the survey and 41 replied. 33 said they
> would keep using the bikes if the scheme continues. 29 named "no bike
> available in the morning" as the biggest problem. 12 asked for helmets
> at the docking station. 6 wrote in the free-text box that the docking
> station is too far from the east entrance. I haven't turned any of this
> into percentages.

## Round one: delegate and check

Send the brief with only the survey note pasted underneath. Here is the
kind of result that comes back.

#### Response

````text
```text
Riders

- 33 of 41 replies would keep using the bikes if the scheme continues,
  and 8 of 41 replies would not.
- 29 of 41 replies named "no bike available in the morning" as the
  biggest problem.
- 12 of 41 replies asked for helmets at the docking station.
- 6 of 41 replies said in the free-text box that the docking station is
  too far from the east entrance.
```
````

It reads well, and it looks like what you asked for. Check it before you
read on. Take the four lines one at a time, with the note next to the
result, and mark each one pass or fail.

#### Checkpoint: Pass or fail, line by line

Check the round one result against the four criteria. Which lines pass and which fail?

Statements:

- A heading and four bullets, each count written as "n of 41 replies"
- Every count is in the note
- The section has no percentages
- No commentary or recommendation

Options:

- Pass
- Fail

The failing line is the second one. The note has no 8 in it. The agent
took 41 replies, subtracted the 33 who said they would keep using the
bikes, and wrote the difference as a finding. It is also a claim the note
doesn't make. The other 8 replies didn't say they would keep using the
bikes, and that isn't the same as saying they would stop. Some of them
may not have answered that question at all.

Notice how the check found it. A read of the whole section gives a
feeling that it is fine. A walk down the four lines gives three passes
and one fail, and the fail names the exact bullet to look at.

## Reading the miss as feedback on the brief

The easy reading of the miss is that the agent made a mistake. The more
useful reading is that the limit in the brief was looser than its check.
The check says every count is in the note, and the brief states that rule
only at the end, as something to tick. Its limit, "Use only the survey
note as your source", allows more, because a number computed from the
note still uses only the note. When two lines of a brief can be read two
ways, the agent may take either reading.

So the fix goes in the brief, as a limit that says what the check tests.
Add this line to the limits of the riders brief:

```text
Copy each count as the note gives it. Don't add, subtract or combine
counts.
```

Then re-send the riders brief with the new line, in a fresh conversation,
and tick the four lines again. This is round two. The first bullet now
ends after "if the scheme continues". All four lines pass. The riders
section is done.

Working this way, one piece at a time with the brief changed after each
check, is iteration. Each result is feedback on the brief, and the next
brief includes what the result showed. The AI Fluency course from Claude
Academy describes the same back and forth between describing a task and
judging what comes back, and it treats the judging as the source of the
next description (AI Fluency: Framework and foundations, Claude Academy).

The cost piece needs the change too, because it has the same gap. The
vendor email has two monthly prices and a twelve-month term, and an agent
with no limit on computed numbers may add a yearly total or the
difference between the two options. Neither number is in the email, and
the manager reads it as if the vendor had quoted it. Put the change into
the cost brief before you send it.

#### Checkpoint: Where does the fix go?

The riders section came back with a count the note never gives. The
agent worked it out by subtracting two numbers in the note. You are about
to delegate the cost piece. Where does the fix go?

- Reply in this conversation that counts are copied, never worked out, and ask for the riders section again.
- Send the riders brief again, unchanged, in a fresh conversation, since the next result may be different.
- Delete the computed count from the bullet by hand, and send the cost brief as you had written it.
- Add a limit on computed numbers to the riders brief, and put the same limit in the cost brief before sending it.

## The cost piece with the change

Here is the cost brief, a model answer to the other half of the previous
exercise, with the new limit. Open a fresh conversation, send it with
only the vendor email pasted underneath, and check what comes back.

#### Prompt (illustrative, not a recorded transcript)

```text
Write the "Cost" section of a report from the vendor email below. Format:
the heading "Cost", then one bullet per option with its monthly price and
what it includes, then one line with the minimum term.

Use only the email as your source. Include both options. Copy each amount
as the email gives it. Don't add, subtract or multiply amounts. Don't add
commentary or recommendations.

Done when: both prices appear once and match the email, each option says
what it includes, the twelve-month term is stated, and every amount is in
the email.

[vendor email follows]
```

#### Response

````text
```text
Cost

- Option A: 1,800 per month for the bikes and the docking station. Your
  facilities team does the maintenance.
- Option B: 2,400 per month, including maintenance and replacement of
  damaged bikes.

Both options have a minimum term of twelve months. Both prices exclude
tax.
```
````

Walk the four lines again. Both prices appear once, and they match the
email. Each option says what it includes. The term is stated. Every amount
in the section is in the email, and there is no yearly total and no
difference between the two prices. All four pass. The section has no computed
numbers, and the log records the change that went in before it ran.

## A two-line log per round

Keep a log while you work, two lines per round: what changed in the brief,
and what the result showed. After the three rounds above it reads like
this.

```text
Round 1, riders piece
  Changed: nothing, this is the first brief from the split.
  Showed: "8 of 41 replies would not" is a count the note never
    gives. The limit was looser than the check, and allowed a
    worked-out count.

Round 2, riders piece again
  Changed: added "Copy each count as the note gives it. Don't add,
    subtract or combine counts."
  Showed: all four lines pass. The first bullet ends after "continues".

Round 3, cost piece
  Changed: added "Copy each amount as the email gives it. Don't add,
    subtract or multiply amounts."
  Showed: all four lines pass. No yearly total or price difference.
```

Read the log and look at what the miss cost. It cost one section of four
bullets, a minute of checking and one re-send. The fix was in the next
brief before the next piece ran. Now picture the same gap in the one-prompt
report from the previous lesson. The computed 8 would sit in a paragraph
next to a yearly total, and both would read as findings. You would find
them after the manager had asked about them, if at all. A short round
keeps each mistake small because it is found in one piece, while the
piece is still the only thing that depends on it.

The log is also the lightest record of what you checked. It has one line
for what changed, and one for what the check found. A record that is light
enough gets kept on a busy day, and the log is light enough for that. The
safety course makes the same point about a checked-by line in a team
template ([Marking what a person
checked](https://schubergphilis.github.io/ai-training/safety/marking-what-a-person-checked/#put-the-mark-in-the-process)).

#### Pitfall: Sending the same brief again

A learner sees the computed count in the riders section, opens a fresh
conversation and sends the same brief again, expecting a better result.
The new result has "and 8 of 41 replies did not say so" in the first
bullet, which is the same computed count with new wording. The brief is
the same, so the loose limit that allowed the miss is still there, and
the agent reads it the same way. A second run of an unchanged brief only
asks for another guess.
When a round misses, change the brief first, and write the change in the
log.

## The round as the unit of work

A round is one brief, one result, one check, and one change to the brief.
The previous lesson made the pieces small enough to check. In this lesson
you worked in rounds, and each round ended with a log entry. A piece
often takes one or two rounds. The next lesson is about the piece that
keeps needing more, and how to see from the log when it is time to do the
step yourself.

Improving work step by step from its failures is also how DeepLearning.AI
teaches builders to improve an agent workflow. They look at where the
output goes wrong, and they fix that step first (Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents, DeepLearning.AI). You do the
same at a smaller size, one piece and one brief at a time.

## Exercise

Delegate the riders piece with the brief above, or with the first piece of
your own split and its criteria. Check the result against the criteria,
one line at a time, and write pass or fail next to each line. Change the
brief with what the result showed, re-send the riders piece with the
change in a fresh conversation, and tick its lines again. Then put the
same change into the brief for the cost piece, and delegate the cost
piece in a fresh conversation. Keep a two-line log per round, in the
format above. It takes about ten minutes. A good result is a log with a
pass or fail for every criterion in each round, and a cost brief that
holds the new limit. If your first round passed every line,
skip the re-send, write "Changed: nothing" for the cost piece and say
what the result showed anyway.
Which line of your log would you want to read before you start the next
report of this kind?

Stretch: Run the usage piece from the previous lesson again with the new limit in its brief, and word the limit so it still allows the one total the brief asks for. Add the round to your log. Did your wording hold?

## Recap

1. Delegate one piece at a time, and check each result line by line against its criteria, pass or fail per line. A read of the whole result gives a feeling, and the lines give you the failing bullet.
2. Read a miss as feedback on the brief. A miss often shows a limit that was looser than the check, such as a limit that allowed a count worked out from other counts. Change the description of the task with what the judging of the result showed (AI Fluency: Framework and foundations, Claude Academy).
3. The fix goes into the brief as a limit, and into the brief of every later piece with the same gap. A reply in the chat stays in that conversation.
4. Keep a log of two lines per round: what changed in the brief, and what the result showed. It is light enough to keep on a busy day.
5. Short rounds keep mistakes cheap, because a miss is found in one small piece before anything else depends on it.

#### Habit

The next time a delegated result misses one of its criteria, write the change into the brief before you send anything else to the agent.

## References

- AI Fluency: Framework and foundations, Claude Academy, https://academy.claude.com/courses/ai-fluency-framework-foundations
- Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents, DeepLearning.AI, https://www.deeplearning.ai/courses/agentic-ai