Agent Unit Economics
Agent unit economics measures whether an agent produces valuable completed outcomes at an acceptable total cost, not whether one model call is cheap.
Source Tweet Artifacts
Core idea
The unit is not "token." The unit is successful outcome.
For an agent, cost per successful outcome includes:
- model input and output tokens;
- cached-token read/write behavior;
- tool/API calls;
- sandbox or machine time;
- browser sessions;
- queues and retries;
- human review time;
- failed runs;
- observability/storage;
- support and recovery.
The useful metric is:
cost_per_successful_outcome =
total_run_cost / successful_outcomes
Then compare that to the value of the outcome and the alternative human/manual workflow.
Why token price is misleading
Agents spend across multiple meters. A cheap model with many retries can be more expensive than an expensive model that finishes. A multi-agent topology can look wasteful on raw tokens but win if it improves success rate or reduces human review. A cached long prompt can change a "too expensive" architecture into a viable one.
OpenAI's prompt caching docs explicitly frame cache hits as reducing latency and cost for repeated long prefixes, while Anthropic's docs expose automatic and explicit caching modes with time-to-live behavior. That means context layout and request routing are economic decisions, not only prompt-engineering details. Source: OpenAI prompt caching, https://developers.openai.com/api/docs/guides/prompt-caching, 2026-06-19; Anthropic prompt caching, https://platform.claude.com/docs/en/build-with-claude/prompt-caching, 2026-06-19
The Coinbase model-spend bookmark is a useful but X-only external observation: Gergely Orosz read Coinbase's reported spend drop as the combination of cheaper inference routes (GLM 5.2/Kimi-class models), smart routing, and caching. Treat it as a hypothesis for an economic playbook, not a confirmed Coinbase primary-source case. The enriched record for the row also attached an unrelated Pollen article, so the article is explicitly excluded as evidence. Source: X/@GergelyOrosz and enriched-record review, 2026-07-03
Vercel AI Gateway provider sorting is the production-grade version of the same routing lever: providerOptions.gateway.sort can rank providers behind a model by cost, TTFT, or TPS at request time. Use it only with evals, latency/error measurement, data-retention checks, spend caps, and outcome-quality tracking; cheaper or faster provider order can still lose if the workflow retries, degrades answer quality, or violates data policy. Source: Vercel changelog/docs, 2026-07-04; Source: Vercel AI SDK
The enterprise-budget paradox is that high AI spend and underuse can coexist. Rahul's bookmark names the failure mode: larger orgs can institute budgets for AI while still not routing enough work through it, because the frontier premium, cheaper-model substitutes, and internal controls are being managed as invoice lines instead of outcome throughput. For Kevin's systems, this reinforces measuring useful work completed per dollar, not "did the org buy AI seats" or "did the org cap tokens." Source: X/@rahulgs, 2026-06-19
Cathy Di's "why are agents not making money rn" row is useful as the market version of the same invariant: an agent product is not monetizable just because the underlying model is capable. The paid unit has to be a complete workflow outcome with a buyer, an error/review boundary, and a gross-margin story after model calls, sandbox time, retries, support, and human review are counted. Otherwise the product sells novelty while the operator eats the actual workflow cost. Source: X/@itsCathyDi, 2026-03-06
Ivan Burazin's app-layer critique adds the differentiation side of the same economics. If the product is only nice UI, a model/data harness, and sales motion, gross margin and retention depend on whether the workflow outcome is proprietary enough to resist cheaper clones. Model optimization is not the only moat, but the paid unit still needs a defensible workflow, data boundary, evaluation loop, or distribution channel; otherwise the sales team is monetizing a commodity harness. Source: X/@ivanburazin, 2026-06-24
Metric stack
| Metric | Why it matters |
|---|---|
| success rate | Cost only makes sense relative to completed work. |
| retry rate | Retries multiply every downstream meter. |
| tool-call count | Tool calls add latency, auth risk, and failure surface. |
| model mix | Route easy work to cheaper/faster models; reserve frontier models for judgment. |
| cached-token ratio | Stable prefixes should become cheaper over repeated turns. |
| sandbox minutes | Long-running agents rent machines, not only models. |
| human minutes | HITL is valuable, but not free. |
| p95 wall time | Slow success can still be unusable. |
| artifact reuse | A run that updates memory or creates reusable tooling has compounding value. |
Economic model
| Work type | Economic shape |
|---|---|
| One-shot answer | model cost + retrieval + citation time |
| Code fix | model/tool cost + tests + review + failed attempt risk |
| Research brief | search/API cost + synthesis + source verification |
| UI implementation | model cost + browser verification + design review |
| Persistent agent | machine uptime + model/tool usage + memory/storage + monitoring |
| Scheduled automation | cron frequency x average run cost x failure/retry rate |
The "agent as service worker" thesis only closes if the service worker's total cost is below the value of the labor it replaces or amplifies. Services-as-Software is a unit-economics claim before it is a product slogan.
Cost levers
| Lever | What it changes | Risk |
|---|---|---|
| Model routing | cost, latency, quality | cheap model may fail silently |
| Prompt caching | repeated context cost/latency | brittle if context order churns |
| Retrieval quality | fewer irrelevant tokens | missed context can cause wrong work |
| Tool batching | fewer round trips | harder error isolation |
| Parallelism | wall-clock time | higher spend and coordination failures |
| Durable state | fewer restarts | persistence and cleanup complexity |
| Human gates | lower risk | higher human time and idle wait |
| Better skills | fewer retries | skill rot if not maintained |
Cheap Tokens Are Not Automatically Cheap Outcomes
A June 2026 bookmark linked a Hacker News discussion about discounted Claude-token resale. Treat that as a market-structure warning, not a procurement recommendation. Grey-market tokens can make a model call look cheaper while adding provenance, account-ban, reliability, billing, and data-handling risk. Source: X/@GregKamradt, 2026-06-25; Source: Hacker News thread, 2026-06-25
The local screenshot makes the economic mechanism explicit enough to preserve: a below-list token price can be subsidized by account pooling, fraud exposure, unstable access, or secondary monetization of outputs. That means "token discount" can move cost out of the invoice and into security review, support burden, data risk, and provider ban risk.
First-party price cuts are different, but still not sufficient. DeepSeek API is a legitimate provider-side example: V4-Pro's post-discount pricing is low enough to change routing candidates, yet the routing question remains cost per successful outcome after evals, latency, retry rate, data policy, spend caps, and provider abstraction. Cheap official output tokens still lose if the workflow retries or fails more often. Source: DeepSeek pricing, 2026-07-03
For Kevin's systems, token price only matters after source legitimacy and operational risk are settled:
| Temptation | Real economic question |
|---|---|
| Buy cheaper tokens | Are they legitimate, auditable, stable, and allowed by the provider terms? |
| Route through a reseller | Who owns failure, refunds, logs, and data exposure? |
| Optimize invoice cost | Does success rate, latency, support burden, or security risk get worse? |
| Ignore spend controls | Can one compromised key or runaway agent drain budget? |
| Ignore resale mechanics | Is the discount funded by a risk that later becomes Kevin's incident? |
Kevin-stack defaults
For Kevin's agent systems:
- track cost per run where product or automation economics matter;
- report p95 and failure rate, not only happy-path demos;
- treat prompt caching as a context-layout constraint;
- route models by task difficulty, not prestige;
- avoid multi-agent fanout unless the task benefits from parallel search, review, or specialization;
- hibernate persistent machines when idle;
- count human review as a cost, even when Kevin is the reviewer;
- value memory updates because they reduce future cost.
IRL comparison
Real providers are already exposing the levers this concept depends on:
- OpenAI exposes cached token counts and cache-hit behavior in usage surfaces.
- Anthropic supports automatic caching for growing multi-turn conversations and explicit breakpoints for stable blocks.
- Sandboxed agent products charge in some combination of runtime, memory, storage, and API usage.
- Observability tools turn traces into cost and latency analysis.
- Nous Portal Teams exposes the organizational layer: shared credit pools, a billing owner, per-member usage visibility, and spend caps. That is the product version of this page's thesis: cost control belongs where teams run agents, not only in a monthly invoice. Source: X/@NousResearch, 2026-06-17
The market is not moving toward "one giant agent call." It is moving toward budgeted workflows with model routing, caching, durable runtime state, and observable run economics.
Failure modes
- optimizing token cost while increasing failed runs;
- measuring average latency while p95 is where users quit;
- ignoring human review time because it is not on the provider invoice;
- using parallelism to hide slow design instead of fixing the workflow;
- caching stale context;
- selling an autonomous service without knowing gross margin per completed task.
Timeline
- 2026-06-19 | Created to connect prompt caching, latency budgets, multi-agent topology, persistent machines, and services-as-software into one cost-per-outcome model. Source: User request; OpenAI and Anthropic prompt caching docs; local Agent Machines economics pages, 2026-06-19
- 2026-06-25 | Added grey-market token economics warning: cheap model access can increase provenance, reliability, billing, and security risk. Source: X/@GregKamradt, 2026-06-25; Hacker News thread, 2026-06-25
- 2026-06-30 | Added Nous Portal Teams as provider evidence for shared credit pools, billing ownership, per-member usage visibility, and spend caps. Source: X/@NousResearch, 2026-06-17
- 2026-07-02 | Added the reviewed local screenshot interpretation: resale discounts can externalize cost into incident, support, and data-governance risk. Source: wiki/assets/x-bookmarks/2069968506951725419/image-01.jpg
- 2026-07-03 | Added DeepSeek API as the first-party counterexample to grey-market token resale: official price cuts can be economically relevant, but still require evals, latency checks, data review, and spend controls before routing agent work. Source: DeepSeek API docs, 2026-07-03
- 2026-07-03 | Added the Coinbase model-spend bookmark as an X-only routing/caching hypothesis and recorded the source-integrity correction: the enriched Pragmatic Engineer/Pollen article attached to that row is unrelated and should not support the model-routing claim. Source: X/@GergelyOrosz, 2026-06-27
- 2026-07-04 | Added Vercel AI Gateway provider sorting as a production routing lever that still requires evals, latency/error checks, data-retention review, and spend caps. Source: Vercel changelog/docs, 2026-07-04
- 2026-07-04 | Added Rahul's enterprise AI-spend/underuse bookmark as the budget-governance failure mode: AI spend can be high while useful routed work remains low, so measure cost per successful outcome instead of raw usage caps. Source: X/@rahulgs, 2026-06-19
- 2026-07-04 | Added Cathy Di's agent-monetization prompt as the market-side failure mode: agent products need a paid outcome unit and margin after retries, sandbox time, support, and review, not only impressive model capability. Source: X/@itsCathyDi, 2026-03-06
- 2026-07-04 | Added Ivan Burazin's app-layer critique as the commodity-harness risk: UI plus model/data access plus sales can sell, but unit economics weaken without a defensible workflow, data boundary, eval loop, or distribution edge. Source: X/@ivanburazin, 2026-06-24