Never re-read what you have already read
An agent that starts every session by reading your whole outline is paying for the same knowledge again and again. The number from this product's own logs: reading the tree cost 962,608 bytes. Asking what had changed since last time cost 1,060. Same knowledge, nine hundred times cheaper, and faster by the same order.
That gap is not a Pando trick. It is the difference between a snapshot and a feed, and most agent integrations only have the first.
Read once, then ask what changed
Every write in Pando is one atomic mutation with a sequence number. An agent that remembers the last number it saw can ask a single question: what happened after this? The answer is the changes themselves, not the tree they happened in.
The habit is three lines in an agent's own memory:
- read the branch once, and write down the sequence number you reached;
- next session, ask for changes since that number;
- write down the new number.
An agent doing this arrives knowing what moved while it was away, which is usually the only thing that matters.
The other three cheap habits
Ask for the least that answers the question. An outline mode drops the notes and the stamps and keeps the shape, and it is about 80% smaller. If the question is "where does this live", the shape is the whole answer.
Write many things in one call. Eleven separate edits came back as 648 bytes when they were sent as one batch, against 3,911 when each echoed its whole bullet. The same eleven edits.
Change part of a line, not the line. Appending to a note without resending it cost 83 bytes against 13,018 for the rewrite. And a rewrite is more than expensive: it is how an agent deletes a sentence you wrote while it was composing.
Why this belongs in the agent's memory, not in your prompt
These are habits, not instructions, and habits go where the agent will find them without you. Written into its memory branch, they survive every session and every context reset. Written into a prompt, they last until you forget to paste it.
The test of whether it worked is simple. Ask your agent, at the start of a session, what it knows. If the answer arrives quickly and starts with what changed rather than with everything, the habits landed.
Try it on your own tree
Connect an agent at pando.ink, let it read one branch, and then ask it tomorrow what changed since yesterday. The first session is the expensive one. Every session after that should not be.