The Harness Effect: How Orchestration Design Slashes Enterprise Agent Token Costs

Agent Cost Composition

Enterprise agents rarely finish a task with a single model call; they involve multiple prompt rounds, tool invocations, retrievals, state updates, and history replay. Each model call generates new input and output tokens, so total cost is the sum of tokens across all rounds.

Core Experiment: Swapping Harness Only

The authors kept 22 tasks, six base models (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6), the evaluation method and price table constant, and replaced only the orchestration layer (the Harness). One run used a traditional production Agent loop as baseline; the second run used the Writer Agent Harness.

Results: the Harness reduced token consumption per task from 14.2 k to 8.8 k, a 38% drop, while quality scores changed only modestly (0.78 → 0.81, not statistically significant given n=22).

How Harness Saves Tokens

Four design choices drive the savings:

Whether to replay the full history or compress it into checkpoints.

Whether tool schemas are injected every round or dynamically narrowed to task needs.

Whether retrieval content is inserted in full or limited to key evidence and citations.

How failure retries are handled—continuous calls versus bounded retries and circuit‑breakers.

These choices affect context length and token usage in each round.

Prompt Structure

The Harness splits prompts into a stable byte‑prefix (tool schema, fixed system prompts, persistent task records) and a volatile tail (time, file status, plans, reminders, voice settings). The stable prefix can be cached; the paper reports 7,886 prompt tokens with 7,876 served from cache (99.9% hit rate).

History Management

Instead of naively appending the entire history each round, the Harness stores persistent execution logs and checkpoint summaries, keeping only recent raw messages. This prevents repeated token consumption and loss of critical decisions when the context window fills.

Cleaning Intermediate Results

Large tool outputs, web pages, or file reads are off‑loaded to the file system or external state; only summaries, previews, references, or pointers are kept in the model context. Sub‑agents perform isolated searches and return concise summaries, avoiding costly full‑context passes.

Optimization Converges to Model Capability

Cost reductions (33%–61%) and latency improvements were consistent across all models, but quality gains depended on model strength. Stronger models (e.g., Palmyra X6, Claude Sonnet 4.6) showed positive quality leverage (+0.079, +0.073), while weaker models sometimes regressed.

Sub‑Agent Capability Threshold

Sub‑agent delegation scores were highest for Palmyra X6 (0.86) and Claude Sonnet 4.6 (0.85), lower for Gemini 3.1 (0.70) and the smaller models (≈0.5). The paper notes that Harness features must be tiered to model ability—limiting tool catalogs or disabling sub‑agents for weaker models.

From Single Tasks to Enterprise Ledger

The authors estimate that for an organization running 1 M agent tasks per month, baseline cost would be $210 k (≈$0.21 per task). With Harness, cost drops to $120 k (≈$0.12 per task), saving $90 k monthly (~$1.08 M annually). They recommend tracking per‑task token accounting (total tokens, input/output split, cache hit rate, schema size, retrieval length, compression effect, retry cost, sub‑agent token isolation) at the Harness layer.

Conclusion

The study demonstrates that orchestration design—not model price—dominates token economics for enterprise agents. By managing context composition, history compression, tool exposure, and retry policies, Harness achieves substantial cost savings while quality improvements remain tied to underlying model capability.