Full history
Pass the accumulated conversation forward. It preserves the record, but grows the amount of context workers receive.
Hashing
is not
retrieval.
A three-arm study of context accumulation in small-model code cartography.
Read the paper · PDFA smaller history.
A clearer handoff.
When one agent hands work to another, what should travel with it? The entire conversation, a compressed identifier, or a retrievable artifact?
This study isolates how those choices change the context sent to workers. Structured artifact passing reduced peak context by 38.3% compared with full-history passing in the controlled evaluation.True within a stated scopeFrozen evaluation report38.3% peak-context reduction against full-history passing, measured inside a two-worker code-cartography harness. It is a result about context cost in that apparatus, not a general claim about reasoning quality.
This is a result from a specific two-worker code-cartography harness. It does not establish a general improvement in reasoning quality.
Peak context · study comparison
Full history
Artifact passing
Less context.
Same question.
Pass the work three ways.
One orchestrator hands a job to a worker. The only thing that changes is what travels with it.
What crosses the wire
TaskTurn 1Turn 2Turn 3Turn 4Turn 5Turn 6
Every previous turn, verbatim
Peak context sent to the worker
Baseline
The record survives intact, and so does everything the worker did not need. Context grows with the conversation rather than with the task.
Normalized to the full-history baseline, measured inside a two-worker code-cartography harness. A result about context cost in that apparatus — not a general claim about reasoning quality.
Three ways to pass the work.
Opaque compression
Pass a compressed representation. A short identifier alone cannot make the underlying state recoverable.
Structured artifacts
Pass inspectable outputs with explicit retrieval. Measure the context actually sent to the worker, then compare it with the baseline.
What this
proves. And
what it doesn't.
- Verified in sourceA three-arm design with a preserved, re-runnable apparatusVerified in sourceStudy repositoryThree-arm design with a preserved apparatus and a reproducibility contract, so the measurement can be re-run rather than believed.
- True within a stated scopeA 38.3% peak-context reductionTrue within a stated scopeFrozen evaluation report38.3% peak-context reduction against full-history passing, measured inside a two-worker code-cartography harness. It is a result about context cost in that apparatus, not a general claim about reasoning quality.
What I'd
test next.
Longer agent chains and broader task classes. The useful question is how well the handoff contract survives different workloads, and where its costs start to show.
Inspect the full study