Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows
Vivek Kumar Singh, Preeti Priyam, Gautam Bhowmick
cs.AI · cs.MA · cs.PF
Quick verdict
Introduces TCA, an exact two-pass attribution of memory-injection cost in multi-agent LLM workflows; credible and candid, but modest effect, single provider.
Summary
This paper introduces the Total Cost of Agency (TCA), a decomposition of multi-agent LLM workflow cost into base prompt, inference, memory injection, miss penalty, and context-accumulation terms, together with an exact attribution method: a two-pass, non-billable token count that measures injected memory tokens with the provider's own tokenizer instead of word-count proxies. On a balanced 200-task enterprise benchmark spanning five categories and workflow depths two to six, run against Claude Haiku and Sonnet APIs, memory injection accounts for 13.6 percent of optimizer-actionable variable cost and a derived 11.8 percent of full billed cost; its share rises from 8.4 percent at depth two to 27.6 percent at depth six, while injected tokens grow linearly with depth (R² = 0.9974). Cutting the retrieval window from 32 to 2 entries reduces injected tokens by 28.7 percent at fixed tier, with accuracy movement within seed-level noise. The authors candidly report that their graph-rewriting transforms are cost-neutral, that two decomposition terms are zero by construction, that tier assignment dominates total cost, and that prompt caching, the most plausible mitigation, is unmeasured. Key limitations are the single provider and framework, synthetic data, single-seed ablations, small absolute dollar stakes, and no released code or data.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Memory-injection cost is a real, separately attributable component of multi-agent workflow bills that production observability tools do not isolate, and it is exactly measurable via a two-pass, non-billable tokenizer count (Sections III-IV).
- On the 200-task benchmark, injection is 13.6% of optimizer-actionable variable cost and a derived ~11.8% of full billed cost, with share rising from a structural zero at depth one to 27.6% at depth six.
- Injected tokens grow linearly with depth over depths two to six (R² = 0.9974, slope ~150 tokens/level), consistent with capacity-bounded theory; the data show no convex growth in the measured range.
- Injection volume is controllable at fixed tier: cutting retrieval window capacity from 32 to 2 entries lowers injected tokens 28.7% with accuracy change within seed-level noise (Table IV).
- The authors' graph-rewriting transforms are cost-neutral, two decomposition terms are zero by construction, total cost is dominated by tier assignment, and prompt caching remains unmeasured (Sections VII-VIII).
Strengths
Unusually disciplined claim scoping
The 'What this article does not claim' paragraph and Section VII ('What did not work') pre-empt over-reading: Condition H's dollar saving is explicitly not attributed to injection (price-vs-volume confound, Section VI-A), conditions within a 1% band are called indistinguishable, and the depth-one zero is labeled structural and excluded from fits
Exact, cheap, and validated measurement method
Eq. (2) counts injected tokens with the provider's own tokenizer via two prompt assemblies; the instrumentation is validated by 1,112 extraction-node observations recording exactly zero injected tokens, matching the structural prediction
Balanced design with deterministic grading
Theory-experiment alignment
Weaknesses
Small absolute effect, partially derived headline
Single-seed evidence at decision-relevant granularity
Single provider, single framework, synthetic data
No code or data release
The decisive mitigation is unmeasured
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Total Cost of Agency (TCA) in LLM workflows?
It is a cost decomposition for multi-agent LLM workflows proposed by this paper: total cost is split into base prompt, inference, memory injection, miss penalty, and context-accumulation terms. Its key insight is that tokens retrieved from memory and injected into each agent's prompt are billed as input tokens but are not reported separately by standard observability tools.
How much of a multi-agent LLM bill comes from memory injection?
On the paper's 200-task enterprise benchmark with Claude Haiku and Sonnet, memory injection was 13.6% of optimizer-actionable variable cost and a derived estimate of 11.8% of the full billed cost. Its share grew from 8.4% at workflow depth two to 27.6% at depth six; at depth one it is structurally zero.
Does prompt caching eliminate memory injection cost?
The paper argues it would not, but did not measure it. Injected memory content is generated within the current task (first injection is always a cache write), nodes have different system prompts that invalidate prefixes, and naive caching can raise cost on writes that are never read. Providers discount cache reads to about 0.1x of input price, so measuring real hit rates is the authors' stated first follow-up.
Is the paper's code or benchmark data available?
No. The paper describes a 200-task synthetic enterprise benchmark, three SQLite databases, and a two-pass accounting harness, but states no repository or artifact release, and the experiments ran against paid Anthropic APIs. Independent verification currently requires reimplementing the harness.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are available only via captions in the extracted text; plotted curve details could not be independently verified
- Table values were transcribed from text extraction; minor transcription risk cannot be excluded
- Text appears complete through the reference list, but section VIII limitations rely on author self-report
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.