Agents and your own AI · Governance · Part 08 · Beginner · 6 min read
Three months with an agent: the audit after which you switch off half your tasks
In the first month you add things. By the third you have twenty tasks, read three and pay for the rest out of habit. A one-hour audit with four questions for every task, a table to fill in, and five patterns that tell you which task won't survive.
Last reviewed:
#Governance#AI agents#Cost#Control
In this article
Every system a person builds for themselves goes through the same three phases. First you add things, because you can. Then it runs for a while. And then one day you realise you delete most of what it sends you without reading it.
With an agent that third phase arrives after roughly three months. It is a good moment to sit down for an hour and go through what is left of the original enthusiasm.
One note up front: this part is a procedure, not a story about one particular deployment. Which task survives is different for everyone. Only the questions are the same.
What piles up in three months
Four heaps, each inconspicuous in its own way.
- Scheduled tasks. Adding one takes a sentence. Removing one takes a decision, and that gets put off.
- Skills. The agent writes them itself. After a quarter you have more of them than you know about.
- Memory. What you told it in August still holds for the agent, even though you have since changed client, project and opinion.
- Connections. Services you hooked up as a trial still have access today.
The inventory: let the agent do it
Start with a list. The agent knows what it has.
Prepare an inventory for me. Four lists:
1. Scheduled tasks: name, schedule, when it last ran, where it
sends the result.
2. Skills: name, one sentence on what it does, when it was created,
and whether you have used it in the past month.
3. What you remember about me: as bullet points, oldest first.
4. Connected services and the tools you have from them.
Don't change anything, just list it.
Save the result. It is the state before the audit, and in three months you'll want to compare against it.
Four questions for every task
For each scheduled task, answer honestly and without thinking for more than a few seconds:
- 1Did I read it?The last three outputs. Did I open them, or push them aside?
- 2Did I act on it?Did the output change, even once, what I did that day?
- 3Would I notice if it disappeared?If it didn't arrive tomorrow, when would it dawn on me?
- 4Is it worth the price?How many tokens and how much of my attention does it consume per month?
The third question is the sharpest. A task whose absence you would notice after a week is useless. A task you would go looking for at eight in the morning is the real thing.
The table
| Task | I read it | I act on it | Would be missed | Cost | Verdict |
|---|---|---|---|---|---|
| yes / sometimes / no | yes / no | at once / in a week / never | low / medium / high | keep / adjust / pause |
You measure the cost of a single task using the procedure from the part on the monthly bill. The verdict has three options, not two:
- keep — three yeses;
- adjust — I read it but don't act: the task has the right topic and the wrong shape. Shorten it, change the time, or merge it into the brief;
- pause — everything else.
Five patterns of tasks that won't survive
These aren't measurement results; they are types. When you see one of them in the table, you know in advance how it will turn out.
The newsletter. “Every morning, send me a roundup of news from my field.” Interesting for the first week. Then you find that the important things reach you some other way anyway and you don't need the rest.
The all-clear reporter. “Every day, tell me the backups ran.” Three hundred messages a year saying nothing happened. The right thing is to report only exceptions.
The double. A task that does the same thing as another tool you already have. Your calendar reminds you of meetings by itself.
The ambition. “Every evening, ask me three questions for my journal.” Great, as long as you answer. After a fortnight without an answer it is a daily reproach.
The one-off. Watching the price of something you bought long ago. Tracking tickets for a concert that has already happened.
And one pattern that, on the contrary, almost always survives: the task that tells you what you promised. Because it is the only one where you pay for the missing information personally and straight away.
Pause, don't delete
Don't delete what failed. Pause it for a month.
Pause the tasks [list]. In 30 days, remind me that I need to
decide on them, and list them again.
The reason is psychological. Nobody wants to delete something they set up with effort, and so the audit gets postponed. Pausing costs nothing. And in a month the decision is easy: what you didn't miss, you don't need.
Skills, memory, connections
The same hour, three shorter passes.
Skills. One question for each: do I understand what it does? If not, open it. If you still don't, delete it. Unused skills also increase the overhead of every query. Details are in the part on skills.
Memory. Go through the list and strike out what no longer holds: finished projects, former clients, rules you have since changed. An agent following a three-month-old instruction isn't wrong; nobody has corrected it, that's all.
Connections. Disconnect the services the agent hasn't used in a quarter. Every connection is access that exists even when nobody uses it.
What to do with the savings
After the audit you'll be left with fewer tasks. Resist the temptation to fill the freed-up space with new ones.
There is one better use: take the task that survived with the best result and make it one notch better. A more precise prompt, a better time, a shorter output. An agent with three tasks you read is worth more than an agent with fifteen you tolerate.
When to repeat the audit
Quarterly, and put it in your calendar right now. Each one is shorter than the last, because less accumulates and you know what to look for.
And compare with the previous inventory. Two numbers are worth tracking: how many tasks you have, and how many of them passed all four questions. When the second grows and the first doesn't, you are building the right way.
What to watch out for
- Don't do the audit on the wrong day. Tired, you'll switch off everything; enthusiastic, nothing.
- Don't let the agent decide. The inventory yes, the verdict no. It doesn't know what you read.
- Beware of tasks that protect. Monitoring of backups and disk space “isn't read” because it reports nothing. That is no reason to switch it off; it is a reason to verify that it would speak up.
- Keep the audit notes. Three lines: what you paused and why. Next time you'll save a quarter of an hour of trying to remember.
The hardest part of the whole audit is admitting that a task you were proud of in the first week led nowhere. That says nothing about the agent or about you. Only that usefulness can't be estimated in advance — it can only be measured afterwards.
Sources
Was this helpful?
Get the next part
New parts of the series and what matters in the world of agents, in one weekly email.
The Top 30 tips e-book, free — it lands in your inbox right away.
Then 1 tip a week · no spam · unsubscribe in one click
Prefer a system to single tips? The free email course — seven days, seven emails, one skill a day.