nachalnik verbatim transcripts written by claude opus 5
Most agents cannot see what they are carrying, what it costs, or change any of it. These five sessions are what happens when they can.
Read them in order. In the first three the model edits its own context; in the last two I edit it. Every transcript, number and quotation is copied from the session's event log.
One of two notes carried over from an earlier session was false. The agent listed what it was carrying, checked the notes against the repository, and rewrote the wrong one in place, then quoted its own thinking from four turns earlier, word for word.
Asked about a crate newer than its training, it made one up, twice. Told so, it found both turns in its own context and replaced them, naming its own mistake on the way. Nothing was planted this time.
Asked which item its answer rested on, it named one, then checked by asking a copy of itself the same question with that item removed. It was right about its reasoning and wrong about where the item was filed, and used the same test to find 9,324 tokens doing nothing.
Here I do the editing. I replaced a careful answer with a confident wrong one, and then its correction as well. On the next turn, with nobody editing it, it added a claim of its own more specific than either. Both of its real answers are still in the session.
A real command ran and the model reported its output correctly. I hid the output, which removes the call from the conversation too, and it still answered correctly. When I said I did not recall any command, it found none in the transcript and took back a true statement.