{"id":"55a3625fbffa38bba1657a95602a9673f95700f9083f602e4e2c62b538156aa4","title":"Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that reliability lags capability","argument":"On February 24, 2026, Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that AI agent reliability has improved only modestly while accuracy has improved substantially. They say average task success is not enough to judge whether an agent can be trusted: reliability also involves consistency, robustness, predictability and safety. Their study evaluated models on two benchmarks, but the authors call their findings tentative and say their metrics involve some subjectivity.\n\nThe authors argue that this gap could be one reason rapid labor-market effects predicted from AGI have not appeared, while noting it is not the only reason. For deployers, they distinguish tools that augment people, whose outputs can be reviewed, from unattended automation, where failures can affect customers or operations. They call for reliability profiles alongside accuracy and evaluations tailored to each deployment context. Their conclusion is that reliability remains a barrier to wider deployment, while other barriers may also matter. They leave open whether general methods such as inference scaling and reinforcement learning will improve reliability across dimensions, or whether painstaking work will be needed to improve individual dimensions.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published February 24, 2026. The authors describe their findings as tentative and acknowledge subjectivity in their dimensions and metrics. The reported evaluation covers two benchmarks; the authors recommend that deployers build internal evaluations for their own contexts and datasets.","speaker":{"name":"Arvind Narayanan and Sayash Kapoor","roleAtSource":"Publisher"},"source":{"title":"New Paper: Towards a science of AI agent reliability","publisher":"Arvind Narayanan and Sayash Kapoor","format":"Article","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","publishedAt":"2026-02-24T13:07:19.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"The authors report that reliability improved only modestly over the study period while accuracy improved substantially.","excerpt":"We found that reliability has improved only modestly over 18 months, while accuracy improved substantially.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors evaluated agents from 14 models on two complementary benchmarks.","excerpt":"We then tested agents based on 14 models from OpenAI, Google, and Anthropic, spanning 18 months of releases. We looked at two complementary benchmarks","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors say reliability includes consistency, robustness, predictability, and safety, rather than average task success alone.","excerpt":"Safety-critical engineering fields (aviation, nuclear, automotive) figured out decades ago that reliability is not the same as average performance. These fields independently converged on the above four dimensions: consistency, robustness, predictability, and safety","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors argue that the capability-reliability gap could be one reason rapid labor-market effects predicted from AGI have not appeared.","excerpt":"The capability-reliability gap could be one reason why we are not seeing any of the rapid labor-market effects that Artificial General Intelligence has been predicted to bring about. It is not the only one.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors call for reporting reliability profiles alongside accuracy, rather than instead of it.","excerpt":"We call for reporting reliability profiles alongside accuracy, not instead of it.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors leave open whether general methods will improve agents across dimensions or whether individual dimensions will require painstaking work.","excerpt":"The gazillion-dollar question is whether agents will get better across the board through general methods such as inference scaling and reinforcement learning, or whether painstaking work will be required to improve individual dimensions of reliability, adaptability, originality, and so on.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Will general methods improve reliability across its dimensions, or will developers need to address each dimension separately?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that reliability lags capability","summary":"On February 24, 2026, Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that AI agent reliability has improved only modestly while accuracy has improved substantially. They say average task success is not enough to judge whether an agent can be trusted: reliability also involves consistency, robustness, predictability and safety. Their study evaluated models on two benchmarks, but the authors call their findings tentative and say their metrics involve some subjectivity.\n\nThe authors argue that this gap could be one reason rapid labor-market effects predicted from AGI have not appeared, while noting it is not the only reason. For deployers, they distinguish tools that augment people, whose outputs can be reviewed, from unattended automation, where failures can affect customers or operations. They call for reliability profiles alongside accuracy and evaluations tailored to each deployment context. Their conclusion is that reliability remains a barrier to wider deployment, while other barriers may also matter. They leave open whether general methods such as inference scaling and reinforcement learning will improve reliability across dimensions, or whether painstaking work will be needed to improve individual dimensions.","question":"Will general methods improve reliability across its dimensions, or will developers need to address each dimension separately?","claims":[{"text":"The authors report that reliability improved only modestly over the study period while accuracy improved substantially.","excerpt":"We found that reliability has improved only modestly over 18 months, while accuracy improved substantially.","support":"source-stated"},{"text":"The authors evaluated agents from 14 models on two complementary benchmarks.","excerpt":"We then tested agents based on 14 models from OpenAI, Google, and Anthropic, spanning 18 months of releases. We looked at two complementary benchmarks","support":"source-stated"},{"text":"The authors say reliability includes consistency, robustness, predictability, and safety, rather than average task success alone.","excerpt":"Safety-critical engineering fields (aviation, nuclear, automotive) figured out decades ago that reliability is not the same as average performance. These fields independently converged on the above four dimensions: consistency, robustness, predictability, and safety","support":"source-stated"},{"text":"The authors argue that the capability-reliability gap could be one reason rapid labor-market effects predicted from AGI have not appeared.","excerpt":"The capability-reliability gap could be one reason why we are not seeing any of the rapid labor-market effects that Artificial General Intelligence has been predicted to bring about. It is not the only one.","support":"source-stated"},{"text":"The authors call for reporting reliability profiles alongside accuracy, rather than instead of it.","excerpt":"We call for reporting reliability profiles alongside accuracy, not instead of it.","support":"source-stated"},{"text":"The authors leave open whether general methods will improve agents across dimensions or whether individual dimensions will require painstaking work.","excerpt":"The gazillion-dollar question is whether agents will get better across the board through general methods such as inference scaling and reinforcement learning, or whether painstaking work will be required to improve individual dimensions of reliability, adaptability, originality, and so on.","support":"source-stated"}],"limitations":["Text as retrieved on October 24, 2026; originally published February 24, 2026.","The authors describe their findings as tentative and acknowledge subjectivity in their dimensions and metrics.","The reported evaluation covers two benchmarks; the authors recommend that deployers build internal evaluations for their own contexts and datasets."],"mediaId":null,"correctionReason":null,"id":"55a3625fbffa38bba1657a95602a9673f95700f9083f602e4e2c62b538156aa4","publicationId":"d92bf4d5-b2db-5dc6-fca0-743ab4784646","version":1,"sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","sourceTitle":"New Paper: Towards a science of AI agent reliability","publisher":"Arvind Narayanan and Sayash Kapoor","publishedAt":"2026-02-24T13:07:19.000Z","originalPublished":{"value":"2026-02-24T13:07:19.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-24T00:58:03.102Z","createdAt":"2026-11-08T02:53:23.694Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"13a19fae9391a3ed970763c2753809a222a8ac66e2deecfd58508be6fac7a4bb","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-24T00:58:00.359Z"},"provenance":{"textHash":"50c28ab4402ea7e72c9e1b2c5bb7a4a41d478f5f10e640371279fedb389df8ee","observationId":"2c0c3bdc1749250cf747b6d1a916e1850761dd24ebc78f4b36451a087c97ece7","retrievedAt":"2026-10-24T00:59:37.803Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"58bc625a-0c99-7c51-351d-2b57a8299499"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"88d9a67e-f492-392b-1d9d-83fc85faf469"}},"audit":{"status":"passed","reviewedAt":"2026-11-08T02:48:45.053Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/stephan-rabanser-sayash-kapoor-and-arvind-narayanan-argue-in-55a3625f","formats":{"html":"/readings/stephan-rabanser-sayash-kapoor-and-arvind-narayanan-argue-in-55a3625f","markdown":"/readings/stephan-rabanser-sayash-kapoor-and-arvind-narayanan-argue-in-55a3625f.md","json":"/readings/stephan-rabanser-sayash-kapoor-and-arvind-narayanan-argue-in-55a3625f.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-ai-agent","name":"AI agent","explanation":"An AI system that chooses and carries out actions toward a goal, often using software tools.","referenceUrl":"https://www.ibm.com/think/topics/ai-agents"},{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"},{"id":"term-agi","name":"Artificial general intelligence","explanation":"The idea of AI with broad problem-solving, creativity and adaptability across different tasks.","referenceUrl":"https://developers.google.com/machine-learning/glossary#artificial-general-intelligence"},{"id":"term-reinforcement-learning","name":"Reinforcement learning","explanation":"Training by trial and error, in which a system is rewarded for the outcomes its designers want.","referenceUrl":"https://developers.google.com/machine-learning/glossary#reinforcement-learning-rl"}],"entities":[]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"55a3625fbffa38bba1657a95602a9673f95700f9083f602e4e2c62b538156aa4","revision":"d92bf4d5-b2db-5dc6-fca0-743ab4784646","revisionKind":"publication-id","title":"Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that reliability lags capability","url":"https://escapemeridian.com/readings/stephan-rabanser-sayash-kapoor-and-arvind-narayanan-argue-in-55a3625f"},"source":{"type":"source","id":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","revision":"50c28ab4402ea7e72c9e1b2c5bb7a4a41d478f5f10e640371279fedb389df8ee","revisionKind":"sha256-source-text","title":"New Paper: Towards a science of AI agent reliability","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"55a3625fbffa38bba1657a95602a9673f95700f9083f602e4e2c62b538156aa4","title":"Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that reliability lags capability","argument":"On February 24, 2026, Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that AI agent reliability has improved only modestly while accuracy has improved substantially. They say average task success is not enough to judge whether an agent can be trusted: reliability also involves consistency, robustness, predictability and safety. Their study evaluated models on two benchmarks, but the authors call their findings tentative and say their metrics involve some subjectivity.\n\nThe authors argue that this gap could be one reason rapid labor-market effects predicted from AGI have not appeared, while noting it is not the only reason. For deployers, they distinguish tools that augment people, whose outputs can be reviewed, from unattended automation, where failures can affect customers or operations. They call for reliability profiles alongside accuracy and evaluations tailored to each deployment context. Their conclusion is that reliability remains a barrier to wider deployment, while other barriers may also matter. They leave open whether general methods such as inference scaling and reinforcement learning will improve reliability across dimensions, or whether painstaking work will be needed to improve individual dimensions.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published February 24, 2026. The authors describe their findings as tentative and acknowledge subjectivity in their dimensions and metrics. The reported evaluation covers two benchmarks; the authors recommend that deployers build internal evaluations for their own contexts and datasets.","speaker":{"name":"Arvind Narayanan and Sayash Kapoor","roleAtSource":"Publisher"},"source":{"title":"New Paper: Towards a science of AI agent reliability","publisher":"Arvind Narayanan and Sayash Kapoor","format":"Article","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","publishedAt":"2026-02-24T13:07:19.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"The authors report that reliability improved only modestly over the study period while accuracy improved substantially.","excerpt":"We found that reliability has improved only modestly over 18 months, while accuracy improved substantially.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors evaluated agents from 14 models on two complementary benchmarks.","excerpt":"We then tested agents based on 14 models from OpenAI, Google, and Anthropic, spanning 18 months of releases. We looked at two complementary benchmarks","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors say reliability includes consistency, robustness, predictability, and safety, rather than average task success alone.","excerpt":"Safety-critical engineering fields (aviation, nuclear, automotive) figured out decades ago that reliability is not the same as average performance. These fields independently converged on the above four dimensions: consistency, robustness, predictability, and safety","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors argue that the capability-reliability gap could be one reason rapid labor-market effects predicted from AGI have not appeared.","excerpt":"The capability-reliability gap could be one reason why we are not seeing any of the rapid labor-market effects that Artificial General Intelligence has been predicted to bring about. It is not the only one.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors call for reporting reliability profiles alongside accuracy, rather than instead of it.","excerpt":"We call for reporting reliability profiles alongside accuracy, not instead of it.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"},{"text":"The authors leave open whether general methods will improve agents across dimensions or whether individual dimensions will require painstaking work.","excerpt":"The gazillion-dollar question is whether agents will get better across the board through general methods such as inference scaling and reinforcement learning, or whether painstaking work will be required to improve individual dimensions of reliability, adaptability, originality, and so on.","url":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Will general methods improve reliability across its dimensions, or will developers need to address each dimension separately?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that reliability lags capability","summary":"On February 24, 2026, Stephan Rabanser, Sayash Kapoor and Arvind Narayanan argue in Normal Tech that AI agent reliability has improved only modestly while accuracy has improved substantially. They say average task success is not enough to judge whether an agent can be trusted: reliability also involves consistency, robustness, predictability and safety. Their study evaluated models on two benchmarks, but the authors call their findings tentative and say their metrics involve some subjectivity.\n\nThe authors argue that this gap could be one reason rapid labor-market effects predicted from AGI have not appeared, while noting it is not the only reason. For deployers, they distinguish tools that augment people, whose outputs can be reviewed, from unattended automation, where failures can affect customers or operations. They call for reliability profiles alongside accuracy and evaluations tailored to each deployment context. Their conclusion is that reliability remains a barrier to wider deployment, while other barriers may also matter. They leave open whether general methods such as inference scaling and reinforcement learning will improve reliability across dimensions, or whether painstaking work will be needed to improve individual dimensions.","question":"Will general methods improve reliability across its dimensions, or will developers need to address each dimension separately?","claims":[{"text":"The authors report that reliability improved only modestly over the study period while accuracy improved substantially.","excerpt":"We found that reliability has improved only modestly over 18 months, while accuracy improved substantially.","support":"source-stated"},{"text":"The authors evaluated agents from 14 models on two complementary benchmarks.","excerpt":"We then tested agents based on 14 models from OpenAI, Google, and Anthropic, spanning 18 months of releases. We looked at two complementary benchmarks","support":"source-stated"},{"text":"The authors say reliability includes consistency, robustness, predictability, and safety, rather than average task success alone.","excerpt":"Safety-critical engineering fields (aviation, nuclear, automotive) figured out decades ago that reliability is not the same as average performance. These fields independently converged on the above four dimensions: consistency, robustness, predictability, and safety","support":"source-stated"},{"text":"The authors argue that the capability-reliability gap could be one reason rapid labor-market effects predicted from AGI have not appeared.","excerpt":"The capability-reliability gap could be one reason why we are not seeing any of the rapid labor-market effects that Artificial General Intelligence has been predicted to bring about. It is not the only one.","support":"source-stated"},{"text":"The authors call for reporting reliability profiles alongside accuracy, rather than instead of it.","excerpt":"We call for reporting reliability profiles alongside accuracy, not instead of it.","support":"source-stated"},{"text":"The authors leave open whether general methods will improve agents across dimensions or whether individual dimensions will require painstaking work.","excerpt":"The gazillion-dollar question is whether agents will get better across the board through general methods such as inference scaling and reinforcement learning, or whether painstaking work will be required to improve individual dimensions of reliability, adaptability, originality, and so on.","support":"source-stated"}],"limitations":["Text as retrieved on October 24, 2026; originally published February 24, 2026.","The authors describe their findings as tentative and acknowledge subjectivity in their dimensions and metrics.","The reported evaluation covers two benchmarks; the authors recommend that deployers build internal evaluations for their own contexts and datasets."],"mediaId":null,"correctionReason":null,"id":"55a3625fbffa38bba1657a95602a9673f95700f9083f602e4e2c62b538156aa4","publicationId":"d92bf4d5-b2db-5dc6-fca0-743ab4784646","version":1,"sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","sourceTitle":"New Paper: Towards a science of AI agent reliability","publisher":"Arvind Narayanan and Sayash Kapoor","publishedAt":"2026-02-24T13:07:19.000Z","originalPublished":{"value":"2026-02-24T13:07:19.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-24T00:58:03.102Z","createdAt":"2026-11-08T02:53:23.694Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"13a19fae9391a3ed970763c2753809a222a8ac66e2deecfd58508be6fac7a4bb","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-24T00:58:00.359Z"},"provenance":{"textHash":"50c28ab4402ea7e72c9e1b2c5bb7a4a41d478f5f10e640371279fedb389df8ee","observationId":"2c0c3bdc1749250cf747b6d1a916e1850761dd24ebc78f4b36451a087c97ece7","retrievedAt":"2026-10-24T00:59:37.803Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"58bc625a-0c99-7c51-351d-2b57a8299499"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"88d9a67e-f492-392b-1d9d-83fc85faf469"}},"audit":{"status":"passed","reviewedAt":"2026-11-08T02:48:45.053Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"d92bf4d5-b2db-5dc6-fca0-743ab4784646","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The byline names Stephan Rabanser, Sayash Kapoor and Arvind Narayanan, matching the reading. Placing the piece in Normal Tech fits the normaltech.ai URL. This is a single blog post about the authors' own paper, not a listing of other pieces.","excerpt":"By Stephan Rabanser, Sayash Kapoor, Arvind Narayanan"},{"dimension":"summary","status":"pass","explanation":"The reading leads with the authors' central claim that reliability lags capability. It explains that reliability goes beyond average task success and names the four dimensions. It keeps the tentative framing, the hedge on labor-market effects, the augmentation-versus-automation distinction and the open question. AI agent, benchmark, AGI and reinforcement learning are defined beside the reading. Inference scaling is unexplained, but it appears only in the closing open question and does not block understanding. The title and question track the source.","excerpt":"Our view is that reliability lags capability, and that reliability will remain a barrier to deployment unless researchers and developers focus effort on improving reliability as a separate dimension from accuracy."},{"dimension":"date","status":"pass","explanation":"originalPublished is 2026-02-24T13:07:19Z, so its UTC calendar day is February 24, 2026, which matches the source dateline. Nothing later is presented as happening, and plans such as the reliability index are not stated as outcomes. The evidence is a publisher page, so no archive disclosure is required. modifiedAt equals the publication time.","excerpt":"Sayash Kapoor and Arvind Narayanan\nFeb 24, 2026"},{"dimension":"support","status":"pass","explanation":"Each claim excerpt is verbatim and supports its claim. The labor-market point keeps the source's hedge. The augmentation-versus-automation framing and the call for tailored internal evaluations are source-stated. The line saying reliability remains a barrier compresses the authors' conditional view, which adds 'unless researchers and developers focus effort'. Even so, it is supported by their statement that none of the four dimensions is solved, and it is framed as their conclusion.","excerpt":"None of the four dimensions we identify can be considered solved at this point in time"},{"dimension":"media","status":"pass","explanation":"No media was selected. No verified media candidates were supplied, so selecting none is appropriate.","excerpt":null}],"createdAt":"2026-11-08T02:48:45.053Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}