SYNTHESIS NOTE
Topics›Work Application Use Cases›this note

Do persistent agents really cost less per token?

When AI agents reuse cached context across tasks, does the standard cost-per-token metric still reveal true economic efficiency? A case study suggests the answer may be no.

Synthesis note · 2026-05-28 · sourced from Work Application Use Cases

A 115-day case study of one physician-scientist running a persistent agentic research environment found that 82.9% of recorded May tokens were cache reads. The workflow was cache-dominant: the agent increasingly reasoned over reused accumulated context rather than fresh inference. The author's inference is that persistent agentic environments may shift the economic unit from cost per token to cost per completed artifact.

This matters because cost-per-token is the native pricing and benchmarking unit, and it systematically misleads about persistent agents. When most tokens are cheap cache reads against a durable memory layer, the marginal token tells you almost nothing about the cost of getting useful work done — the expensive resource is the accumulated context and reusable procedures that make each new task cheap. Two agents with identical token counts can differ enormously in artifacts produced.

The counterpoint is that cost-per-artifact is hard to standardize — "artifact" is fuzzy (a paragraph? a paper? a repository?) and reproducible artifact-level denominators barely exist, which is exactly why the field defaults to tokens. But defaulting to the measurable wrong unit is still wrong. Therefore the methodological recommendation that follows is concrete: future evaluations should adopt artifact-level denominators and cost-per-artifact estimates, because the economics of a stateful, cache-dominant agent live at the artifact level, not the token level.

Inquiring lines that read this note 62

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

What happens to knowledge when intelligence becomes tokenized like a commodity? Why do token-level mechanisms matter for learning to reason? What do systematic disagreements between annotators reveal about ground truth? How does misalignment propagate through agent communication networks? How do multi-agent LLM systems fail distinctly compared to single agents? How can AI chatbots provide therapeutic benefit without causing harm? How do surface patterns enable correct outputs but reduce robustness? Do language models reason like humans or mimic surface patterns? How does AI adoption across firms reshape employment and inequality? When do multi-agent systems provide sufficient quality returns on token investment? How should test-time compute scaling work in agentic systems? How should inference compute be allocated based on problem difficulty? How do prompting refinements mask underlying biases and model frequency patterns? What structural properties of attention create systematic model biases? How do standardized protocols improve multi-agent coordination and reliability? Why do standard benchmarks fail to predict agent deployment success? How should agent systems validate and persist generated code artifacts? What should agent evaluation prioritize to reveal reliable behavior? Can compression size predict model complexity better than parameter count alone? How should agents manage memory granularity to improve long-term performance? When do multi-agent systems outperform single frontier models? How do capability benchmark scores systematically misrepresent true model abilities? What attack surfaces do reasoning traces and chains introduce? How vulnerable are token issuance and authorization policies to coordinated attacks? Can memory architectures handle ultra-long context better than attention? How does harness optimization generalize across different model architectures and domains? How effective are honeytokens and decoys against different security threats?

Related concepts in this collection 4

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 115 in 2-hop network ·medium cluster Open in graph ↗
Do persistent agents really cost less per token? Why does agent efficiency differ from model size r… How should we measure agent system performance bey… What happens to code that agents create and then s… Will agents compete for attention just like users …

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Should artifact-level benchmarks replace token counts for agent evaluation?How do cache-dominant workflows change the marginal cost of agent tasks?What makes a standardized artifact unit measurable across different research domains?Can two agents with identical token counts produce vastly different outputs?
Original note title

persistent agentic environments shift the economic unit from cost per token to cost per completed artifact