$ cat paper/abstract.md

Agents that work beyond a single context window, or in teams, hand their work to a successor as a short written summary. Evaluation has focused on whether an agent can solve a task, and relatively little is known about how much of what an agent has learned survives being written down. We study this in 56 interactive worlds with hidden rules, where an agent that is told nothing fails and an agent given a 300-character statement of the rules succeeds. We compare one agent that plays eight lives and keeps its whole history with a chain of eight agents that each play one life and inherit a 300-character note.

All six models we test learn less in the chain: after eight generations it stands where the single agent stood after about four lives. Controlled variants of the chain decompose this handoff loss into three components: none from changing the agent, a loss from compressing to 300 characters, and, on the worlds where the loss is large, a larger loss from compressing repeatedly. When each note is written from the full record of earlier lives instead of from the previous note, the same 300 characters recover 70% to 83% of the loss for the two strongest models. The loss appears in every model we test and is not alleviated by instructing the writer to record evidence, to revise the inherited note, or to keep a structured ledger. When the screen instead names the rule that ended each life, the loss on the hardest worlds disappears.

In simpler terms, an agent cannot hand over what it has not yet understood: a note holds what its writer has worked out, and a rule that takes several lives to see is lost with the lives that were not passed on.

@misc{ratchet2026,
  title  = {Agents Cannot Hand Over What They Have Not Yet Understood},
  author = {Anonymous},
  year   = {2026},
  note   = {RATCHET benchmark}
}