How to see what an AI agent did on your LinkedIn account
Short answer: read the activity log, and check it against LinkedIn's own Sent folder. Do not rely on the chat transcript. The transcript records what the model believed was happening. The log records what the code actually did.
The gap between the transcript and the truth
An agent runs a batch of twelve connection requests. Three fail on a rate limit. The model receives twelve tool results, reads them quickly, and tells you it sent twelve.
Nothing malicious happened. The model summarized. The transcript now says twelve and the reality is nine, and you will discover the difference in a week when three people you were counting on never replied.
This is why the log has to be written by the code that performed the action. That code knows whether the API call returned a success. It has no summary to make and no reason to round.
What a log should record
| Field | Why it has to be there |
|---|---|
| The exact text | A description of the message is not the message. You need to be able to read what the other person read. |
| Timestamp | Pace is what LinkedIn watches. You cannot see your own pace without times. |
| Success or failure | A silent failure looks exactly like a message nobody answered. |
| Failure reason | Rate limit, expired session, and blocked recipient need different responses from you. |
| The target | So a wrong recipient is visible rather than inferred. |
| Append-only | A record that can be edited afterward is not a record. |
Checking it against LinkedIn
The log is one source. LinkedIn is the other, and it is the one that decides.
- Messaging, Sent. Every direct message that actually landed.
- My Network, then Manage invitations, then Sent. Outstanding connection requests, which is also where you find out whether the volume is higher than you thought.
- Your activity, then Comments. Every comment posted under your name.
- Your activity, then Posts. Everything published as you.
Do this once in your first week with any tool. If the two sources disagree, you have learned something important early rather than late.
Questions people ask next
Can I just ask the agent what it did?
Yes, if the log is exposed to it as a tool. Then the answer is read out of the record rather than recalled from the conversation, and the two are very different things. Without that tool the agent is answering from memory of a chat that may have been summarized or truncated.
How long should the log be kept?
Longer than the profile data. Profile and post content is worth deleting quickly because it is personal data about other people. The record of what you sent is yours, and it is the thing you will want when something goes wrong months later.
What if a tool has no activity log?
Then LinkedIn's Sent folder is your only record, and you have no way to see failures at all. For a tool that reads profiles this is survivable. For a tool that sends messages under your name it is not.
Does an activity log help if my account gets restricted?
It helps you understand why. Restrictions follow volume and repetition, and a log with timestamps shows both. It also tells you which day the pattern changed, which is usually the day something was misconfigured.
What about actions I took by hand?
Those will not be in the tool's log, because the tool did not do them. This is another reason to check LinkedIn directly. The complete picture of your account is the log plus your own activity, and only LinkedIn holds both.
How Iridium does it
Every attempted action is written to an append-only ledger: what ran, when, whether it succeeded, and the exact text that was sent. Failures are recorded with their reasons. Entries cannot be edited or removed after the fact.
You can read it in the Activity screen in the web app. Your agent can read it too, through the get_activity_log and get_last_action tools, so asking "did that actually post" gets an answer from the record instead of from the conversation.
Every action on the record
7-day free trial. No credit card. Cancel anytime.
Start free → Read the FAQ