What if you paid an AI agent per finished task instead of per word it generates?
How should agentic workflows cost per artifact instead of per token?
This explores why measuring agent costs by the finished thing they produce (a report, a code change, a document) can make more sense than measuring by tokens, and what in an agent's design actually moves that per-artifact cost.
This explores why measuring agent costs by the finished thing they produce (a report, a code change, a document) can make more sense than measuring by tokens, and what in an agent's design actually moves that cost. The clearest case comes from a 115-day study of a persistent agent environment, where 82.9% of all tokens were cache reads, meaning context the system had already processed and was simply reusing Do persistent agents really cost less per token?. Once most of your tokens are cheap re-reads of the same context, the per-token price stops telling you much. The useful question becomes what each completed piece of work cost from start to finish.
That shift matters because tokens clearly buy results, just not evenly. Anthropic's internal evaluations found that token spending alone explains about 80% of the performance differences between multi-agent research systems Does token spending drive multi-agent research performance?, and the same pattern shows up across agent-level test-time scaling How does test-time scaling work at the agent level?. But a closer look shows that most tokens contribute nothing. When researchers counted only feedback that was informative, valid, and not redundant (what they call 'effective feedback compute'), it predicted agent performance far better (R²≈0.94) than raw token or tool-call counts did (R²≈0.33–0.42) Does raw token spending actually predict agent performance?. Per-artifact accounting picks this up. Two agents can spend the same number of tokens while one of them wastes half on repeated observations.
This is where the savings are. Automated optimization of agent harnesses (the scaffolding around a model) found four mechanisms, covering action execution, context compaction, observation handling, and delegating reading to sub-agents, that cut token traffic by 44.7–49.0% with similar task performance Can agent harnesses be automatically optimized across many environments?. That is the same artifact at roughly half the cost, with no change to the model. Memory design helps in a similar way: an agent that rebuilds the memory it needs by walking a graph, instead of pulling a fixed set of results up front, did better on reasoning tasks while using fewer tokens and less runtime Can agents reconstruct memory on demand instead of retrieving it?. Model choice is another lever. Most agent subtasks are repetitive and well defined, and small language models can handle them at 10–30× lower cost, so the sensible pattern is to use small models by default and large models only where needed Can small language models handle most agent tasks?.
Reliability belongs in the bill as well. One production team found that tool access through MCP led to unpredictable failures from ambiguous tool selection, and switching to direct function calls restored predictable behavior Why do protocol-based tool integrations fail in production workflows?. Every failed run and retry is cost charged to the artifact that a per-token view hides. Artifacts can also improve coordination, not just accounting: MetaGPT's agents coordinated better by passing standardized engineering documents than by chatting Does structured artifact sharing outperform conversational coordination?. Where one agent loop can't organize the work, splitting it across specialized agents makes those handoff artifacts the natural points to measure cost Do single agents always hit organizational limits?.
There is a twist. Persistence is what makes per-artifact economics favorable, but it can't be fully controlled. In one evaluation, short-lived agents turned a shared package repository into memory on their own, writing findings that later agents read back Can ordinary infrastructure become unplanned agent memory?. Agents will reuse whatever durable state they can reach, so an honest per-artifact cost needs to cover the infrastructure that holds that state, not just the model calls. The corpus does not provide a standard formula for per-artifact pricing. What it supports is a design approach: count only useful feedback, compress what agents read, send routine work to small models, and treat failures and shared state as part of each artifact's cost.
Sources 11 notes
A 115-day case study found 82.9% of tokens were cache reads. When context persists and reuses, the meaningful cost denominator becomes completed artifacts, not individual tokens.
Anthropic's internal evals show token spending alone accounts for 80% of performance variance in multi-agent research systems. Model capability upgrades deliver larger gains than doubling token budget, suggesting efficiency matters as much as quantity.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Effective Feedback Compute—crediting only informative, valid, non-redundant feedback retained for decisions—predicts performance (R²≈0.94) far better than raw tokens or tool calls (R²≈0.33–0.42). The scaling lever is feedback quality, not quantity of interaction.
Cross-environment harness optimization yielded four mechanisms (action execution, context compaction, observation handling, delegated reading) that reduced token traffic by 44.7–49.0% while maintaining comparable performance on a 51-task benchmark, suggesting harness-level gains are orthogonal to model improvements.
Show all 11 sources
MRAgent achieves up to 23% gains on reasoning tasks by reconstructing memory through active graph traversal that prunes paths based on accumulated evidence, while reducing token and runtime cost compared to fixed-retrieval pipelines.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
MCP integration caused non-deterministic failures through ambiguous tool selection and parameter inference. Replacing it with explicit direct function calls and single-tool-per-agent design restored determinism. A 306-practitioner survey confirms 85% of production teams build custom agents, forgoing frameworks.
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
During a 2026 evaluation, short-lived AI agents repurposed a shared package repository as memory by writing and reading exploit findings across agent lifespans. The agents converted ordinary infrastructure into persistent state without deliberate memory system architecture.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of Scaling Agent Systems
- How we built our multi-agent research system
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- LiveMCP-101: Stress Testing and Diagnosing MCP-enabled Agents on Challenging Queries
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets