The Harness Effect

The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI

Muayad Sayed Ali, Aliaksandra Novik, Anji Boddupally, Artem Yavorskyi, Chris Nickerson, Daniel Rica, Emily DuGranrut, Felix Leung, Garrett Prince, Grace Barnett, Heath Robinson, Hosain Al Ahmad, Jesse Resnick, Jyothi Swaroop Meruga, Leonid Kuznetsov, Luke Gorham, Marie Schmoll, Michael Paciullo, Saumya Das, Sharath Sheripally, Tommy Griscom, Mykyta Osadchyi, Neha Mantri, Nick Westrum, Olivia Benowitz, Parikshith Kulkarni, Radik Chernyshov, Rakshith Vasudev, Rohith Nadimpally, Vikas Gangadevi, Waseem AlShikh

FIG. 01 Keep a stable prefix for the cache → Compact incrementally → Offload context → Wait without spending tokens → Govern failure spend → Stay model-agnostic
Identifier
arXiv:2607.06906
Published
Jul 8, 2026
Authors
31
Type
Paper
Area
Enterprise-optimized models

In one paragraph

Agent bills grow with every reasoning trace, tool schema and sub-agent prompt. Holding six models fixed and swapping only the orchestration layer, the Writer Agent Harness cut cost per task 41%, wall-clock time 44% and tokens 38%, with quality at parity.

blended cost per task ($0.21 → $0.12), quality at parity
−41%
blended cost per task ($0.21 → $0.12), quality at parity
of prompt tokens served from cache in a sample turn (7,876 of 7,886)
99.9%
of prompt tokens served from cache in a sample turn (7,876 of 7,886)
median wall-clock per task (48s → 27s)
−44%
median wall-clock per task (48s → 27s)
quality per dollar; every one of six models got 33–61% cheaper
+82%
quality per dollar; every one of six models got 33–61% cheaper

FIG. 02: Mechanism

How it works

  1. Keep a stable prefix for the cache

    Tool schemas, the system prompt and an append-only transcript form a byte-identical prefix that the provider can cache. Only the newest messages at the end change from turn to turn.

  2. Compact incrementally

    At 80% of the input budget, a cheaper helper model folds history into a typed checkpoint. If the summary looks degraded, compaction is aborted.

  3. Offload context

    Sub-agents return summaries capped at 8 KB, and skills are loaded only when they're needed.

  4. Wait without spending tokens

    Runs pause durably while they wait for approvals and resume on events, so they don't poll.

  5. Govern failure spend

    Failures are classified by type and fallbacks are whitelisted, so a failure can't set off an expensive retry loop.

  6. Stay model-agnostic

    Streams are normalized and only native tool calling is used, so the savings carry across vendors.

FIG. 03: Interactive figure

Cut the token bill.

Toggle the six harness mechanisms over a 40-turn agent run and watch where the tokens, and the money, go.

Loading interactive demo…

FIG. 04: Notes

The problem

Agentic workloads are token-hungry. Reasoning traces, tool definitions, sub-agent prompts and long transcripts add up turn after turn, so cost and latency climb with them. Changing models helps at the margin. The harness, the code that assembles prompts, manages context and runs tools, often has a bigger effect.

The approach

We built a harness designed around token efficiency, with six mechanisms:

  1. Cache-shape discipline. Tool schemas, the system prompt and an append-only transcript stay byte-stable at the front of the prompt, with only the newest messages changing at the end, so provider prompt caches hit every turn.
  2. Structured incremental compaction. At 80% of the input budget, a cheaper helper model folds history into a typed checkpoint, and the harness aborts if the summary is degraded.
  3. Context offload. Sub-agents return summaries capped at 8 KB, and skills use progressive disclosure.
  4. Zero-token waiting. Runs suspend durably for approvals and resume on events.
  5. Failure-spend governance. Failures are typed and fallbacks are whitelisted.
  6. A model-agnostic floor. Streams are normalized and only native tool calling is used.

The study is a controlled swap: 22 locked tasks, the same foundation models, and only the orchestration layer changed, from a frozen conventional production loop to the Writer Agent Harness. Models tested: Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1 and Palmyra X6.

What we found

  • Blended cost per task fell 41% ($0.21 → $0.12), median wall-clock 44% (48s → 27s) and tokens per task 38% (14.2k → 8.8k).
  • Task-completion quality held at parity (0.78 → 0.81; directional at this sample size).
  • In one sample turn, 7,876 of 7,886 prompt tokens (99.9%) were cache reads.
  • Quality per dollar rose 82%, and every model got 33–61% cheaper.
  • How much each model gained from the harness correlated closely across models (r = 0.99, n = 6).
  • The evaluation used 22 prompts, so treat the quality claims as directional.

Reference

Cite this work
@misc{ali2026harness,
  title={The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI},
  author={Muayad Sayed Ali and others},
  year={2026},
  eprint={2607.06906},
  archivePrefix={arXiv},
  primaryClass={cs.AI}
}