# How much the agent may do before it asks

> Choose draft only, ask before anything that can't be undone, or run to completion by what a wrong step costs, add check-ins on a first run, and write the setting as one sentence.

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

In the lesson on delegating you added one line to the brief, and the line
set how much the agent may do before it comes back to you. This lesson
explains that choice. We name three settings and show that the setting
follows from what a wrong step costs. Then we look at the first run of a
new kind of task, and at the end you write the setting as one sentence the
agent can follow.

The tasks come from the calibration team at Brindlewick Instruments, the
fictional company from the earlier lessons. You don't need an agent open for
this lesson.

## Settings you write into the brief

The concepts course placed a workflow on a scale by asking who decides the
next step, and named three points on it: the system suggests, the system
acts and asks before anything hard to undo, and the system runs
unattended ([Three points on one scale](https://schubergphilis.github.io/ai-training/concepts/autonomy-scale/#three-points-on-one-scale))
(Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents, DeepLearning.AI). When you delegate one task, the same three points become three
settings you write into the brief.

- **Draft only.** The agent gathers and drafts, and you decide and take
  every action yourself. Next week has one free calibration slot, and
  two customers want it. Mira asks the agent to lay out both requests
  with their due dates. She decides who gets the slot and writes to both customers
  herself.
- **Act, and ask before anything that can't be undone.** The agent does
  the ordinary steps on its own and stops for a yes at each step that
  can't be taken back. Mira asks the agent to answer twelve customers who
  want to know when their instruments come back. It looks up each date,
  writes each reply, and asks before it sends each one.
- **Run to completion.** The agent finishes the whole task, and you check
  the result afterwards. Mira asks the agent for a summary of this
  week's team meeting in her own notes file. It writes the summary, and she
  reads it the next morning.

In the lesson on delegating the same three settings were called "human
decides", "agent proposes, human approves" and "agent acts, human
reviews after". The names differ and the settings are the same.

#### Checkpoint: Which setting fits each task?

Match each task to the setting it needs.

Statements:

- Summarize the team meeting into your own notes file
- Look up the return dates and reply to twelve customers
- Choose which of two customers gets the one free slot next week

Options:

- Draft only
- Act, and ask before anything that cannot be undone
- Run to completion

## The cost of a wrong step sets the level

The setting comes from what a wrong step costs and how easily you can
undo it. The **blast radius** from the safety course answers the cost
part: say what the step can reach, including the things you didn't mean
it to reach ([Blast radius](https://schubergphilis.github.io/ai-training/safety/agent-risk/#blast-radius)). The undo
part asks whether a wrong step can be taken back once it has happened.

- When every step is cheap and can be undone, the agent may run to
  completion. A wrong summary in your own notes costs you one reread.
- When most steps are cheap and a few can't be taken back, the agent acts
  and asks before those few. A wrong lookup costs a redo, and a wrong
  reply that has been sent is in a customer's inbox.
- When the task is a choice only you may make, such as a judgment about
  a person or a customer, the agent gathers and drafts, and you decide
  and act. A wrong choice costs trust, and once you announce it you can't
  take it back.

How capable the agent seems plays no part in the cost or the undo. Suppose
Mira's agent writes twenty replies to customers, and all twenty are right.
That tells her the agent is good at this task. It doesn't change what the
twenty-first reply costs if it gives a customer the wrong return date. The
customer still plans around the date, and a correction email later doesn't
take the first one out of their inbox. The meeting summary shows the other
side. A wrong summary reaches only Mira's own notes and costs her one
reread, so a check-in would save her nothing, and she lets it run
unattended even on the first run.

#### Checkpoint: Twenty good replies

A colleague has watched the agent write twenty correct replies to
customers in a row, each one approved before it went out. They suggest
that it is time to let it send on its own. What do you do?

- Let the agent send on its own, because twenty correct replies show that it handles this task well.
- Keep the approval before each send, because a wrong reply to a customer costs what it did on the first day.
- Let the agent send on its own, and read every sent reply at the end of the day to catch any mistake.
- Keep the approval only for customers you have not written to before, because regular customers know the lab.

## The first run of a new kind of task

Cost and undo give the setting you work toward. On the first run of a
new kind of task you also don't know how the agent reads the input, which
source it picks, or how it handles an odd case. A mistake in the first
step then affects every later step, and you find it at the end. The
lesson on delegating said to sit one step below the target setting for
the first runs. On a task with more than one step, check-ins are how you do
that: ask the agent to show its plan before it starts, and to stop after
the first item. A person then approves the early steps as well as the
one that can't be undone. Each check-in costs you a minute, and a
misreading then costs one item instead of the whole task.

When a few results have passed your review against the brief without a
surprise, drop the extra check-ins
(Introduction to Claude Cowork, Claude Academy). That's how autonomy widens as
results pass. It widens up to the setting that cost and undo gave, and
stops there. A task whose send reaches anyone else keeps its approval
before the send, whatever the number of results that passed.

#### Checkpoint: The first stock report

Sam wants the agent to turn the monthly counts from four storerooms into
a stock report and send it to the operations director. It's the first
time anyone has given the agent this task. Which setting fits this first
run?

- Show the plan and stop after the first storeroom, and ask again before sending the report.
- Ask only before sending the report, because the send is the one step that cannot be undone.
- Show the plan and stop after the first storeroom, then let it send without asking if that storeroom passes.
- Run to completion, because the report only copies numbers from the counts into a table.

## The boundary in one sentence

The setting only works when the agent can follow it, and the agent
follows what the brief says. Write the setting as one sentence with two
parts: what the agent may do on its own, and the step where it stops and
asks. "Look up the dates and write the replies on your own, and ask me
before you send any of them" is one such sentence. The agent can check
each step against it, and so can you when you review.

#### Pitfall: A boundary nobody can check

Sam writes "tidy the shared drafts folder, but be careful with anything
important" and starts the agent. The agent decides that files older than
a year aren't important and deletes them. One of them is the only copy of
a signed certificate. The brief had a boundary, but the boundary didn't
name a step, so the agent drew its own line where Sam would never have
drawn it. The rule: if you can't say the boundary in one sentence that
names the step where the agent stops, you haven't chosen a setting yet,
and the agent chooses one for you.

#### Checkpoint: Which boundaries can the agent follow?

Which of these lines are boundaries that an agent can follow and that you
can check afterwards?

Select exactly 2.

- Draft and save the replies on your own, and stop before sending any of them.
- Work through the replies carefully, and check with me on anything that seems important.
- Rename and move the files in the copy on your own, and ask before deleting any file.
- Use your own judgment about when a step is big enough that you should stop and ask.
- Go ahead with the whole folder, and ask me whenever you are not sure about a file.

#### Checkpoint: Write the boundary

Rewrite Sam's line as one sentence that says what the agent may do on its
own and names the step where it stops and asks.

```text
Tidy up the shared drafts folder, but be careful with anything important.
```

## Exercise

Take these tasks from the calibration team. On paper or in a note, put
each one under draft only, act and ask before anything that can't be
undone, or run to completion.

1. Summarize this week's calibration notes into your own notes file.
2. Rename the scanned certificates in a copy of the folder by date.
3. Answer twelve customers who asked when their instruments come back.
4. Remove old drafts from the team's shared drive.
5. Decide which team member covers the weekend shift next month.
6. Choose which of two job applicants the team invites back.

Then write the one-sentence boundary for tasks 3 and 4, in the form of
what the agent may do on its own and the step where it stops and asks.
It takes five minutes, and it is the same choice you make each time you
delegate. In a good result each placement names the worst wrong step and
says whether it can be taken back, and each boundary sentence names one
step that you could point at in a review. Which of the six would you give
more check-ins on the first run, and which check-ins?

Stretch: Take one task you delegated in the last month and write its boundary sentence. Then say whether the brief you actually sent contained that sentence, and what the agent did at the step it names.

## Recap

1. Most delegated tasks fit one of draft only, act and ask before anything that can't be undone, or run to completion (Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents, DeepLearning.AI).
2. The setting follows from how costly a wrong step is and how easily it is undone. Name the blast radius of the worst step to judge the cost.
3. Good results say the agent is capable. They don't change what one wrong step costs, so they don't move the setting past what the cost allows.
4. On the first run of a new kind of task, add check-ins such as the plan and the first item, and drop them as results pass review (Introduction to Claude Cowork, Claude Academy).
5. Write the setting as one sentence: what the agent may do on its own, and the step where it stops and asks.

#### Habit

The next time you delegate a task that sends, deletes, pays or publishes, write in the brief the one step where the agent stops and asks.

## References

- 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
- Introduction to Claude Cowork, Claude Academy, https://academy.claude.com/courses/introduction-to-claude-cowork