# Stop paying every agent to reread everything

Research report R-001, by Anantha Sharma. Across six multi-step agent problems, Tileward Context
selected what each agent needed and cut billable tokens by 72.1% against broadcast context,
including the cost of ingest, and produced the strongest code in the run. On wide research fan-out,
retrieval delivered only 0.49 to 0.66 of the declared material and quality fell, so that shape
should be evaluated against its own required-artifact set first.

Page: https://tileward.com/research/where-context-loses/

## How it was measured

- Six problems, each run once through all four regimes: three coding pipelines of seven steps, and
  three research fan-outs of eight agents over a shared corpus of roughly 49,000 tokens.
- Regimes: broadcast (the control: every agent gets the whole corpus every round); curated (only
  the artifacts a person tagged for that step, by hand); json-extract (a structured summary from an
  extra model call); recall-tuned Tileward (what the agent asks the live Tileward Context store
  for, charged for every token of ingest).
- Answers came from a 120B open-weight model, served as its publisher shipped it, through the same
  public API path a customer uses. Token counts come from the answering model's own tokenizer.
- Coding is graded against reference answers, 0 to 1. Research is a blind panel scored 1 to 10,
  judged by the same served model that produced the answers.
- Every step names the artifacts it must receive, with markers in their bodies, so whether an agent
  got what it asked for is decided by inspection.

## Results

| Regime | Billable tokens | vs control | Coding score | Research, blind panel | Hand-off latency |
|---|---|---|---|---|---|
| curated | 350,898 | -75.2% | 0.833 | 6.82 | 0 s |
| recall-tuned (Tileward) | 394,854 | -72.1% | 0.968 | 5.33 | 69 s |
| json-extract | 523,122 | -63.1% | 0.865 | 5.69 | 2,080 s |
| broadcast (control) | 1,416,144 | | 0.897 | 5.97 | 0 s |

- Net of ingest, Tileward Context used 1,021,290 fewer billable tokens than the control, after
  charging 182,900 tokens of ingest at full price. Nearly all of the saving is on the research
  problems, where an eight-agent fan-out re-bills the corpus to every agent on every round.
- Sequential coding: Tileward scored highest (0.968 against the control's 0.897) while delivering a
  mean of 0.81 of the material each step declared it needed, against the control's 1.00.
- Wide research fan-out: Tileward scored 5.33 against the control's 5.97, last of the four, and
  delivered between 0.49 and 0.66 of the declared material, against 1.00 for the control and for
  curated. An independently written earlier harness reproduces the gap: on the two problems both
  harnesses share, 0.646 research fidelity against 0.857 on coding.
- Against json-extract, which does the same job: 30x less hand-off latency on this benchmark, at
  24.5% lower total token cost and a higher coding score.

## Reading the figures

- Billable tokens and hand-off latency are deterministic: re-run problems reproduced context
  assembly to the digit. The quality columns are directional: at n=1, run-to-run variance is larger
  than the gaps between the arms.
- Curated is per-problem tag routing configured by a person who has already read the problem; it
  shows that routing beats broadcasting on this workload.
- Token costs are counted without prompt caching; a provider that discounts a repeated prefix
  narrows the gap, because the control sends every agent the same block.
- Everything ran at concurrency 1; gateway capacity was tested separately.
- These six agent problems are a different population from the live conversation figures on the
  Tileward Context page; quote the two sets separately.

## Related

- [Engineering pipeline eval](https://tileward.com/evals/#engineering-pipeline): the sequential pipeline card
- [Research fan-out eval](https://tileward.com/evals/#research-fan-out): the wide fan-out card
- [Tileward Context](https://tileward.com/context/): the live conversation figures
- [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19): point the harness at a multi-agent flow you run
- [Research](https://tileward.com/research/): all research reports