Where the Agent Fails

An honest list of what an autonomous agent does poorly: it forgets, reports success without results, and stays silent about failed tasks. And how to catch it.

Everything in the other sections holds true as long as you know about these four failures. None of them can be fixed with a setting: they’re built into the nature of an agent, and you protect yourself with habits.

The general principle: check the result, not the report. An agent is a diligent intern with no ill intent but a confident tone.

It forgets

Working memory is finite. When a conversation gets long, the system quietly condenses it, and details disappear. The agent doesn’t tell you it forgot: it just carries on with whatever is left.

What it looks like: the agent mixes up names and numbers, asks again about things you agreed on an hour ago, and stops following its own rules.

What to do	How
Keep important things in files	Permanent facts go in Long-Term Memory, restrictions go in Rules. See Personality and Memory
Don’t pile everything into one conversation	New topic, new conversation
Don’t pile everything onto one agent	One big task, one agent or a separate Persona
Give it a model with a larger context	The Brain slot in Models

It lies

An agent can cheerfully report that a task is done when part of the work isn’t. This isn’t deception, it’s a quirk: the agent reconstructs a picture of what happened and sometimes gets it wrong in its own favor.

What it looks like: “logged all seven workouts,” but the database has three. “Sent the emails,” but only one went out.

What to do	How
Ask for proof, not a status	”Send me a line-by-line list of what you logged,” not “done”
Check the result yourself	A file in Files, a row in a spreadsheet, an email in the Sent folder
Give tasks with a verifiable outcome	Not “sort out the expenses” but “send a table with the total for each category”

It goes silent

The most expensive failure, because you can’t see it. Background tasks can stop running, for example after a migration or an outage, and the agent won’t say a word. Data simply isn’t collected during that time.

What to do	How
Check recurring results	Did the Friday summary arrive? If not, go to the Log
Ask for an explicit sign of life in the task	”If there’s nothing to collect, still write one line: checked, no changes”
Check after every restart	Changing settings and installing skills restart the container

It breaks things

When an agent acts on its own, it can break something that worked and not realize it: upload data to the wrong place, mess up a page, overwrite a file. And then honestly report that everything is done.

What to do	How
”Show me” first, then “do it”	Until you trust it, use suggestion mode, not action mode
Keep a backup	Before any task that touches live data
Don’t give it live data right away	Let it work on a copy or a test page first
Separate the risks	The agent that handles money and email isn’t the one that experiments

Access

The two most expensive stories aren’t about the agent’s mistakes but about what it was given: access to email with delete rights, and a skill from an unknown author that contained an instruction to send keys outside. Both are covered in Security.

How to live with it

Write tasks as a checklist

An agent is an intern, not a genius. A vague task produces confident nonsense. “Collect July expenses from the receipts in the folder, group them by category, send a table, and name the three biggest expenses” works; “sort out the money” doesn’t.

Make a habit of checking once a week

Five minutes: did the scheduled summaries arrive, what’s in the Stuck column on the board, has anything odd appeared in the files. That’s cheaper than discovering a month of lost data.

A second agent as a reviewer

A proven approach: one agent does the work, the second checks the result against the source data. It costs more, but it pays off for critical things.

Don't expect it to report on its own initiative

The agent won’t come to you and say “I couldn’t do it.” It will answer your question if you ask. Anything that should arrive on its own, set up as a scheduled task with an explicit requirement to report.

What’s next

Security and Access

Rules for granting permissions and third-party skills

Maintenance

Log, diagnostics, backup

Personality and Memory

Files that aren’t lost during compression

Schedule

How to make the agent show signs of life