--- title: From Tokens to Verifiable Work: Agent Infra Is Rewriting... description: As AI agents evolve from single model calls into long-running, multi-tool execution systems, the industry's current observability stack—call telemetry,... canonical: https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: From Tokens to Verifiable Work: Agent Infra Is Rewriting Its Unit of Measure | daily.dev og:description: As AI agents evolve from single model calls into long-running, multi-tool execution systems, the industry's current observability stack—call telemetry,... og:url: https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9 og:image: https://api.daily.dev/og/posts/AnC9zsGc9.png og:image:alt: From Tokens to Verifiable Work: Agent Infra Is Rewriting Its Unit of Measure og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # From Tokens to Verifiable Work: Agent Infra Is Rewriting Its Unit of Measure **[eunomia eBPF Blog](https://daily.dev/sources/eunomia)** · 11 min read · 0 upvotes · 0 comments ## Summary As AI agents evolve from single model calls into long-running, multi-tool execution systems, the industry's current observability stack—call telemetry, execution traces, delivery activity metrics—fails to answer whether work was actually completed correctly. A cluster of recent releases from OpenAI, Google Cloud, GitHub, and Alibaba Cloud's AgentLoop all point toward outcome-based measurement, but each still relies on proxies. The post proposes a 'verifiable work unit' abstraction that bundles intent, permissions, an evidence chain, a deterministic outcome oracle, total cost (including retries and human review time), and full version attribution. Research like DeepSWE's hand-written behavioral verifiers and AgentSight's eBPF-based system tracing fill in parts of this picture, but no single product yet delivers the complete abstraction. Practical advice: define 'done' as an externally verifiable state, separate activity metrics from outcome metrics, and build replayable failure entries for regression testing. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the difference between an execution trace and a verifiable work unit for AI agents? An execution trace records what an agent did—model calls, tool invocations, latency, token counts—but cannot confirm whether the user's goal was met. A verifiable work unit goes further by bundling a clear 'done' condition, permissions in effect, an evidence chain, a deterministically checkable outcome oracle, total cost including retries and human review, and version information for replay. Traces explain execution; work units judge value. _Teams shipping agent workflows track the gap between trace coverage and real outcome verification on daily.dev._ ### How reliable are AI coding agent benchmarks like SWE-Bench for measuring real task completion? Benchmark reliability is fragile. A study replaying 740 performance-optimization tasks across four cloud machine types found that reference patches passing on one machine fail on another, and 9 of 28 pairwise rankings reversed depending on score aggregation method. DeepSWE's hand-written behavioral verifiers disagreed with an LLM judge only 1.4% of the time, versus 32.4% disagreement for SWE-Bench Pro's inherited tests, suggesting purpose-built verifiers are substantially more reliable. _Developers choosing between agent evaluation frameworks find benchmark comparisons like this on daily.dev._ ### What components should a complete AI agent work unit include for accountability and cost tracking? A complete agent work unit should include: clear intent and 'done' condition; permissions and human-approval boundaries active during execution; an evidence chain of model calls, tool calls, state changes, and output artifacts; a deterministically checkable outcome oracle (test, deployment status, ledger receipt); scoring basis and human-review outcome when deterministic judgment is impossible; total cost covering tokens, runtime, retries, and human wait time; and agent, model, prompt, tool, and environment versions for replay and attribution. _Engineers designing agent accountability systems follow infrastructure patterns like these on daily.dev._ ## Similar posts on daily.dev - [AI-Assisted Engineering: Measure Outcomes, Not Activity](https://daily.dev/posts/ai-assisted-engineering-measure-outcomes-not-activity-kgafc31vm) · Thomas Thornton · 33 upvotes · 5 comments - [Long-running Agents](https://daily.dev/posts/long-running-agents-irtro0veg) · Addy Osmani · 78 upvotes · 3 comments - [Why every AI agent decision needs a receipt](https://daily.dev/posts/why-every-ai-agent-decision-needs-a-receipt-toqfrephk) · The New Stack · 1 upvotes · 0 comments - [Agent observability: measuring tools, plans, and outcomes](https://daily.dev/posts/agent-observability-measuring-tools-plans-and-outcomes-bvc6yqfi8) · portkey · 1 upvotes · 0 comments - [Un-observable AI is Un-trustworthy AI](https://daily.dev/posts/un-observable-ai-is-un-trustworthy-ai-lly8obp7q) · Coralogix · 2 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#observability](https://daily.dev/tags/observability), [#opentelemetry](https://daily.dev/tags/opentelemetry) [View this post on daily.dev](https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"From Tokens to Verifiable Work: Agent Infra Is Rewriting Its Unit of Measure","url":"https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9"},"datePublished":"2026-08-11T08:54:22.062Z","dateModified":"2026-08-11T09:11:20.078Z","description":"As AI agents evolve from single model calls into long-running, multi-tool execution systems, the industry's current observability stack—call telemetry,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b59e04eefe0ea85df0b7e50151b0b12e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b59e04eefe0ea85df0b7e50151b0b12e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"eunomia eBPF Blog","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"eunomia eBPF Blog","logo":"https://media.daily.dev/image/upload/s--z-nTl7TK--/f_auto,q_auto/v1786438304/logos/eunomia?_a=BAMAMicg0","url":"https://daily.dev/sources/eunomia"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,observability,opentelemetry","timeRequired":"PT11M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"eunomia eBPF Blog","item":"https://daily.dev/sources/eunomia"},{"@type":"ListItem","position":3,"name":"From Tokens to Verifiable Work: Agent Infra Is Rewriting Its Unit of Measure"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/from-tokens-to-verifiable-work-agent-infra-is-rewriting-its-unit-of-measure-anc9zsgc9#faq","mainEntity":[{"@type":"Question","name":"What is the difference between an execution trace and a verifiable work unit for AI agents?","acceptedAnswer":{"@type":"Answer","text":"An execution trace records what an agent did—model calls, tool invocations, latency, token counts—but cannot confirm whether the user's goal was met. A verifiable work unit goes further by bundling a clear 'done' condition, permissions in effect, an evidence chain, a deterministically checkable outcome oracle, total cost including retries and human review, and version information for replay. Traces explain execution; work units judge value. Teams shipping agent workflows track the gap between trace coverage and real outcome verification on daily.dev."}},{"@type":"Question","name":"How reliable are AI coding agent benchmarks like SWE-Bench for measuring real task completion?","acceptedAnswer":{"@type":"Answer","text":"Benchmark reliability is fragile. A study replaying 740 performance-optimization tasks across four cloud machine types found that reference patches passing on one machine fail on another, and 9 of 28 pairwise rankings reversed depending on score aggregation method. DeepSWE's hand-written behavioral verifiers disagreed with an LLM judge only 1.4% of the time, versus 32.4% disagreement for SWE-Bench Pro's inherited tests, suggesting purpose-built verifiers are substantially more reliable. Developers choosing between agent evaluation frameworks find benchmark comparisons like this on daily.dev."}},{"@type":"Question","name":"What components should a complete AI agent work unit include for accountability and cost tracking?","acceptedAnswer":{"@type":"Answer","text":"A complete agent work unit should include: clear intent and 'done' condition; permissions and human-approval boundaries active during execution; an evidence chain of model calls, tool calls, state changes, and output artifacts; a deterministically checkable outcome oracle (test, deployment status, ledger receipt); scoring basis and human-review outcome when deterministic judgment is impossible; total cost covering tokens, runtime, retries, and human wait time; and agent, model, prompt, tool, and environment versions for replay and attribution. Engineers designing agent accountability systems follow infrastructure patterns like these on daily.dev."}}]} ```