- The paper demonstrates through a controlled 22-task experiment across six models that changing only the orchestration harness reduces tokens per task by 38%, cost by 43%, and wall-clock time by roughly half while keeping headline quality statistically comparable.
- The paper identifies cache-shape discipline, structured compaction, context offloading, zero-token waiting, and bounded failure recovery as practical mechanisms that make orchestration design the main driver of token economics.
- The paper finds that efficiency gains apply across all tested models, but quality improvements depend on model capability, with weaker models sometimes regressing on orchestration-heavy tasks and delegated sub-agents requiring a capability floor.
This paper by a large team at Writer, Inc. argues that in enterprise agentic AI, the orchestration layer—the "harness"—is the decisive determinant of token economics, and demonstrates this with a controlled experiment that swaps only that layer while holding models, tasks, judges, and prices fixed (2607.06906).
Framing: token maxing and the harness as price-setter
The paper's central diagnostic construct is token maxing: a development trajectory in which token intensity per task grows while marginal quality per token declines—quality purchased at a worsening exchange rate. The authors argue falling per-token prices finance this habit via a Jevons dynamic: efficiency gains lower unit cost and raise total consumption, so total spend rises even as per-token prices fall. Benchmark tables report quality; invoices report tokens; token maxing is invisible to the former and visible only to the latter.
The formal framing decomposes a k-turn agent loop's input tokens into harness-controlled terms: system prompt, replayed history, tool schemas, retrieval payloads, and user turns, multiplied by retries and dead-end branches. A naive full-history-replay harness grows input tokens quadratically in turn count (O(k2)); compaction, prefix caching, and offload convert this to approximately linear. Two further facts sharpen the economics: agent workloads are input-dominated (input-to-output ratios near 100:1), and cached input is billed at roughly 0.1× list price, giving an effective input price pineff​=pin​(1−h(1−κ)) with cache-hit fraction h and κ≈0.1. Since h is a function of prompt byte-stability across turns—a property set entirely by orchestration code—the harness controls both how many tokens are submitted and the price at which the dominant share is billed.
The controlled swap
The experimental design is a paired comparison: 22 locked, capability-audited enterprise tasks spanning nine capability areas (grounding/retrieval, Playbooks, MCP tool use, presentations, voice, image analysis, identity/refusal, content generation, plus net-new sub-agent delegation), executed on six foundation models across five vendors and three weight classes (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6). One arm runs a conventional production agent loop frozen on 2026-06-07—a monolithic ~49 KB system prompt replayed every turn, regex-parsed XML tool calls, destructive middle-truncation, polling-based waits. The other runs the Writer Agent Harness at its intended GA configuration. Cost is computed identically for both arms from trace-shim-recorded token counts against one pinned OpenRouter price table.
The headline result: blended across models, cost per task falls from $0.21 to $0.12 (−43%), median wall-clock roughly halves (a 1.8× speedup), and tokens per task fall from 14,200 to 8,800 (−38%), while headline task-completion moves from 0.78 to 0.81—a wash the authors explicitly decline to claim as improvement at n=22. Derived efficiency metrics move decisively: quality per dollar rises from 3.71 to 6.75 (+82%), and completions per million tokens rise from roughly 55 to 92.
Model invariance of efficiency, capability dependence of quality
Per-model results show no exceptions: cost reductions range from −33% (Gemini 3.1) to −61% (Flash 3.5), with latency reductions from −33% to −55%. The largest relative gains land on the fast tier, consistent with the decomposition: for cheap models, harness overhead is a larger share of the bill. A pointed comparison: under the baseline, switching from the most expensive model (Palmyra X6, $0.25) to the cheapest (Qwen 3.6, $0.16) saves 36%; adopting the harness while keeping any model saves 43%. On this workload, orchestration design moved cost more than the entire model menu did—an inversion of where optimization attention typically goes, and an argument that organizations renting their orchestration layer have outsourced the variable they control most.
Quality behaves differently. Across 48 capability×model cells, 30 improve, 11 are flat, and 7 regress—and all seven regressions occur on the three smaller models, concentrated in orchestration-heavy capabilities (MCP tool use drops Qwen from 0.65 to 0.50). Collapsing each model to its mean quality gain yields "harness leverage": the gain tracks baseline strength almost perfectly (O(k2)0, over six points, which the authors flag as suggestive rather than conclusive). Stronger models convert harness structure into quality; weaker models experience it as load. The economic consequence is asymmetric: efficiency gains are unconditional, quality gains are earned by capability.
The harness's one net-new capability—delegated sub-agents with scoped context and capped summaries—crosses a usable reliability threshold only on the two strongest models (Palmyra X6 at 0.86, Sonnet 4.6 at 0.85), versus 0.42–0.45 on the fast tier. This is a concrete instance of a capability floor: orchestration features are not free to expose, implying graceful degradation by model tier and routing on feature demand rather than prompt difficulty alone.
Mechanisms
Six mechanism families implement the savings, mapped onto the cost decomposition:
- Cache-shape discipline — a two-zone prompt with a byte-stable prefix (tool schemas, stable system prompt, append-only transcript) carrying up to four provider cache breakpoints, and a volatile tail structurally banned from caching. Measured on an identical-prefix call, 99.9% of prompt tokens (7,876 of 7,886) were served as cache reads.
- Structured compaction — typed checkpoints (durable memory, execution summary, verbatim user requirements) at 80% of budget, incremental fold-forward, summarization on a cheaper helper model off the paying loop.
- Context offload — sub-agents as context firewalls returning ≤8 KB summaries with citations on sidecars; progressive skill disclosure; tool outputs spilling to files beyond size caps; event-sourced plan state rendered once per turn.
- Zero-token waiting — durable suspension instead of polling turns, with write-ahead journaling and generation fencing so crashes resume rather than re-buying a run's tokens.
- Failure-spend governance — typed failure classification before fallback, mid-stream failures discarded without side effects, circuit breakers on byte-identical failing calls, bounded loops.
- A model-agnostic floor — typed route plans, normalized provider streams, native tool calling only, schema hygiene for weaker models.
The through-line is that token economy and output quality are one lever: removing stale or bulky tokens simultaneously cuts cost and cleans the model's working set, consistent with long-context degradation findings.
Landscape comparison and fleet economics
An architectural survey of widely used agent systems (Claude Code/Cowork, LangGraph, CrewAI, AutoGen/AG2, Hermes Agent), assessed from documentation and a design-time source study—not head-to-head measurement—finds that per-task token accounting is nowhere a published contract. Without a meter in the spending layer, the authors argue, token maxing is unobservable and therefore unmanaged. Anthropic's own figures (agents at ~4×, multi-agent systems at ~15× chat-level token consumption) are cited as Definition 1 operating as an architecture.
At fleet scale, the blended rates imply an organization running one million agent tasks monthly pays $O(k^2)$1120k under the harness—$1.08M annually from an orchestration change alone. Three properties make this saving unusual: it is model-portable (implemented above the API, applying to future models), volume-linear, and stacks multiplicatively with price declines, routing, and prompt compression rather than substituting for them.
Limitations and open questions
The paper is candid about several constraints. At O(k2)2, quality deltas are directional only; the headline is parity, not improvement. The baseline was run once and frozen, leaving run-to-run variance unmeasured. Task-completion is LLM-judged, with judge bias mitigated but not eliminated. Dollar figures inherit one pinned price table (token and latency results are price-independent). The task set mirrors an enterprise assistant workload and may not transfer to long-horizon coding benchmarks. The harness-leverage correlation rests on six points. Only one baseline–harness pair, both from a single vendor, was measured, so magnitudes are specific to that pair; cross-harness claims are architectural, not benchmarked. Notably, the multi-step research synthesis task regressed in quality (0.80 → 0.60) despite a 46% cost cut—the one place aggregate parity conceals a real trade—and the release posture holds open-weight candidates pending a fix. The disclosure states all authors are Writer employees, and the evaluated harness and Palmyra model are Writer products.
Conclusion
Holding models constant and swapping only the orchestration layer, six foundation models completed the same work for 38% fewer tokens, 43% less money, and nearly half the wall-clock time, with headline quality at parity. Efficiency gains were uniform across every model tested; quality gains scaled with baseline capability; and the one new capability carried a capability floor. The paper's substantive claim is that token intensity is a design choice made in code above the model API, measurable per release, and governable there—and that completions-per-million-tokens belongs alongside quality in agent release gates. The open questions it leaves are empirical: whether the leverage relationship holds over a wider model panel, whether the magnitudes generalize beyond one vendor pair and workload shape, and what configuration-level cross-harness measurement would show.
