When colleagues ask me "should we build an agent for this?", I walk them through the same ladder. One everyday job, turning meeting notes into follow-up tasks, solved at five levels of autonomy. The meeting is invented; effort and risk are my own rough estimates.
01 · The ladder
The same job can be solved five ways.
The job has seven parts, and they stay the same all the way up. What changes is who does each one. Slide up the ladder and watch the parts move from you to software, and what you get at the end.
The machinery behind the higher steps is drawn on other pages: how a tool works in fig. 2 of An agent is a loop with tools, how a skill loads only when it is needed in fig. 3 of What happens when you press Enter, the loop in which an agent decides its next step in fig. 3 of An agent is a loop with tools, and agents handing work on in its fig. 4 and subagents in fig. 5 of What happens when you press Enter.
Step 3 of 5: A skill with tools
The meeting notes
Planning meeting, Tuesday
- Ana sends the three venue options by Friday.
- Budget needs a check before we book. Who?
- Ben calls the caterer about the date.
The job
You, by hand
Software
decides what happens next
Start it
You: you run the skill
Say what to do
Software: the skill says how
Write the tasks
Software: writes the tasks
Find the owners
Software: reads the team list
Put them in the task tool
Software: drafts; you confirm
Chase late ones
You: you send reminders
Friday report
You: you write it
By hand: 3 of 7 parts
What you get
Drafts in the task tool
- Send the venue optionsAna, Friday
- Check the budgetMia, from the team list
- Call the catererBen
Nothing is saved until you confirm the drafts.
Show as text
| Part of the job | 1. A single prompt | 2. A shared template | 3. A skill with tools | 4. An agent | 5. A team of agents |
|---|---|---|---|---|---|
| Start it | You: you open a chat | You: you open the template | You: you run the skill | Software: after every meeting | Software: agent 1, after each meeting |
| Say what to do | You: in your own words | Software: a saved template | Software: the skill says how | Software: the agent says how | Software: one brief per agent |
| Write the tasks | Software: writes a list | Software: one format for all | Software: writes the tasks | Software: writes the tasks | Software: agent 1: summarise |
| Find the owners | You: you fill in names | You: you fill in names | Software: reads the team list | Software: asks the team chat | Software: agent 2: assign |
| Put them in the task tool | You: you copy them over | You: you copy them over | Software: drafts; you confirm | Software: creates them | Software: agent 2 creates them |
| Chase late ones | You: you send reminders | You: you send reminders | You: you send reminders | Software: reminds the owners | Software: agent 3: follow up |
| Friday report | You: you write it | You: you write it | You: you write it | You: you write it | Software: agent 4: report |
| Decides what happens next | You | You | You | Software | Software |
| Parts by hand | 6 of 7 parts | 5 of 7 parts | 3 of 7 parts | 1 of 7 parts | 0 of 7 parts |
| What you get | A list in your chat. You copy it into the task tool. Asked another way, it comes out another way. | A list in the team format. Same columns for everyone who uses the template. You still copy it over. | Drafts in the task tool. Nothing is saved until you confirm the drafts. | Saved in the task tool. Nobody checks before it acts. You read its log afterwards. | Saved, chased and reported. Four agents hand the work on. You read their logs. |
| How it works | You paste the meeting notes into a chat and ask for a task list. | A saved prompt with slots for the meeting type, the team and the notes. Everyone uses the same one. | A packaged skill reads the notes, looks up who is on the team and puts draft tasks in the task tool. | Runs after every meeting on its own: reads the notes, asks owners when something is unclear, creates the tasks and chases late ones. | One agent per stage: summarise, assign, follow up, and report to the manager every Friday. |
Each step up hands more of the job to software. From the agent on, software also decides what happens next.
02 · Gain and watch
A step up costs more, and a mistake gets further.
Building gets more expensive fast: from 15 minutes for a good prompt to about three months for a team of agents. What you have to watch grows with it. Follow one mistake through each setup: as long as a person reads the result before it moves on, the mistake stops there. Without that person it keeps going, and the further it gets, the more people it reaches. Some mistakes only exist higher up: a prompt never sends anyone a reminder.
So every step up needs its own safeguard: an owner for the template, a confirm step and tight permissions for the skill, logs, limits and a way to undo for the agent, and tests and monitoring at every hand-over for a team of agents.
A task for the wrong person: The AI reads "Ben calls the caterer" as a task for Ana.
1A single prompt
15 minutes
Who stops it: You, reading the list
2A shared template
an hour
Who stops it: You, reading the list
3A skill with tools
a day
Who stops it: You, confirming drafts
4An agent
three weeks
Who stops it: Nobody: it reaches Ana
5A team of agents
three months
Who stops it: Nobody: it reaches the manager
1 dot = 1 working day of building; the specks of steps 1 and 2 are circled. A bar on the line is you, checking before the mistake moves on.
Show as text
| Step | Cost to build | A task for the wrong person | A reminder that keeps coming | So you add |
|---|---|---|---|---|
| 1. A single prompt | 15 minutes | Gets as far as the task list. Who stops it: You, reading the list. You read the list before you copy it over. | Cannot happen here. Nothing sends reminders: you do. | Nothing yet: you read every answer. |
| 2. A shared template | an hour | Gets as far as the task list. Who stops it: You, reading the list. You read the list before you copy it over. | Cannot happen here. Nothing sends reminders: you do. | Someone who owns the template. |
| 3. A skill with tools | a day | Gets as far as the task tool. Who stops it: You, confirming drafts. It lands as a draft; you see it when you confirm. | Cannot happen here. The skill stops at drafts. You send the reminders. | A confirm step and tight permissions. |
| 4. An agent | three weeks | Gets as far as a colleague. Who stops it: Nobody: it reaches Ana. Nobody checks before the agent saves it and reminds Ana. She tells the agent it is not hers; you still write the Friday report yourself. | Gets as far as a colleague. Who stops it: Nobody: Ana gets them all. Ana gets ten reminders in one day, until someone adds a limit. | Logs, limits and a way to undo. |
| 5. A team of agents | three months | Gets as far as the manager. Who stops it: Nobody: it reaches the manager. Ana tells the follow-up agent; the assigning agent never hears it, and the Friday report lists Ana as late. | Gets as far as the manager. Who stops it: Nobody: it reaches the manager. Ana gets ten reminders, and the report agent tells the manager she ignored them. | Tests and monitoring at every hand-over. |
The two mistakes: A task for the wrong person: The AI reads "Ben calls the caterer" as a task for Ana. A reminder that keeps coming: A late task gets a reminder every hour instead of once a day.
Every step up adds power and adds things you have to watch.
03 · Your idea
Place your own idea on the ladder.
Think of a job you want AI to help with and answer five questions. Each yes lifts the floor to the step next to it on the ladder; the highest yes decides. The answers start filled in for the meeting-notes job from fig. 1: change any of them for yours.
Your job
Yes toYour step
- 5needs 4
5A team of agentsif yes to 5, with 4
- 4
4An agentif yes to 4
- 23
3A skill with toolsif yes to 2 or 3
- 1
2A shared templateif yes to 1
-
1A single promptwhere everyone starts
Highest yes: questions 2 and 3. It reads or writes other systems: a skill with tools, with a person confirming before anything is saved. Separate stages alone do not need a team: without a yes on question 4 they fit in one setup.
Show the questions and rules as text
Everyone starts at step 1, a single prompt. Each yes lifts the floor; the highest yes decides. Question 5 only counts together with a yes on question 4.
| Question | A yes lifts you to |
|---|---|
| 1. Do several people do this job, the same way, more than once a week? | Step 2: A shared template |
| 2. Does it need data from another system (agenda, CRM, task tool)? | Step 3: A skill with tools |
| 3. Should it write or change things in another system? | Step 3: A skill with tools |
| 4. Should it decide its own next step, with no person checking in between? | Step 4: An agent |
| 5. Are there separate stages that different specialists would handle? | Step 5: A team of agents, only with a yes on question 4 |
Filled in for the meeting-notes job: 1 yes, 2 yes, 3 yes, 4 no, 5 yes. Result: step 3, a skill with tools. Highest yes: questions 2 and 3. It reads or writes other systems: a skill with tools, with a person confirming before anything is saved. Separate stages alone do not need a team: without a yes on question 4 they fit in one setup.
The highest yes decides the step, and nothing more.