# Give agents useful context, not the whole backlog

Tileward research report R-005. Across 104 AgentBench tasks with real earlier work in the history, Tileward Context solved 37 while sending 523,028 context tokens; full histories solved 24 while sending 1,651,372. The 68% (3.2×) context reduction is stable; the 13-task gap is not, because identical runs can flip 14–19 task outcomes, so the report stands behind the repeated ordering rather than a fixed accuracy margin.

Page: https://tileward.com/research/a-third-of-the-context/

## How it was measured

- **Benchmark:** 104 [AgentBench](https://arxiv.org/abs/2308.03688) tasks, scored as solved or not. Each task starts with about 4,100 tokens of earlier work already in the agent's history.
- **What changed:** every approach sees the identical history; the only difference is what that history is turned into before it is sent.
- **Model:** one model, one model family.
- **Repeats:** Tileward Context ranked ahead of full history in two production runs. The accuracy gap tests at p = 0.011 on this run and p = 0.078 on a near-identical run, three tasks apart.

## The result

| What the agent was given | Solved | Context sent | Less than everything |
|---|---|---|---|
| Tileward Context | 37 of 104 | 523,028 | 68% less (3.2×) |
| The newest turns that fit the same budget | 32 of 104 | 596,087 | 64% less (2.8×) |
| Nothing (history deleted) | 29 of 104 | 279,503 | 83% less (5.9×) |
| Everything, as-is | 24 of 104 | 1,651,372 | Baseline |
| The benchmark's own trimming | 0 of 104 | 565,140 | 66% less (2.9×) |

## The controls

- **The same setup, run twice:** a 14–19 task wobble from run-to-run luck alone.
- **Newest-first, same budget:** five tasks behind Tileward Context, on more context, so the result is not just from sending fewer tokens.
- **Delete the history entirely:** eight tasks behind. Deleting the history still beat sending all of it, so part of what compression buys here is removing distraction rather than remembering anything.
- **Measure against production:** a local copy flattered the result; production is the number in the table.

## A benchmark defect

The zero in the table reproduced three times and is the only statistically unambiguous effect in the study. AgentBench trims long histories by keeping the first message plus a recent window. In its operating-system tasks the first message is the protocol and a worked example, and the task statement sits several messages later, so past the trimming threshold the agent keeps the rules and loses the job, while the run completes with plausible-looking low scores. Anyone running long-horizon agent work through that default path is affected.

## Next

If you are evaluating context compression yourself, run the same setup twice before believing any accuracy difference. Read this alongside R-006, where long conversations produced the other result: far less context, but a lower answer rate than the full transcript.

## Related

- [Agent backlog eval card](https://tileward.com/evals/#agent-backlog): the result and limits in one screen
- [R-006](https://tileward.com/research/a-nineteenth-of-the-context/): the long-conversation result
- [AgentBench paper](https://arxiv.org/abs/2308.03688): the benchmark
- [Tileward Context](https://tileward.com/context/): how it works
- [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19): bring an agent workload with real accumulated history
- [Research](https://tileward.com/research/): every report