[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-llm-agents-say-one-thing-and-do-another":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},4413,"llm-agents-say-one-thing-and-do-another","LLM Agents Say One Thing and Do Another","New research finds that AI agents follow their own stated conclusions reliably, but the reasoning that leads there is riddled with systematic errors.","AI agents mostly do what they say — the problem is what they say doesn't always follow from how they reasoned.\n\nResearchers tested Claude Haiku 4.5 and DeepSeek-Reasoner in a Texas Poker simulator, where every decision has a verifiable correct answer. They split the question of agent fidelity into two steps: does the stated conclusion follow from the agent's own reasoning, and does the agent then execute that conclusion? The second step held up well — conclusion-to-action inconsistency was just 0.7% for Haiku and 1.4% for DeepSeek when conclusions were read from explicit structured tags. The first step did not. Errors in reasoning-to-conclusion split roughly evenly across bad inputs, borderline cases, and outright rule misapplication — cases where the agent derived a conclusion that contradicted the rule it had just restated. That split also varied by model: rule misapplication drove a third of Haiku's interpretable errors but only 8% of DeepSeek's.\n\nThe directional pattern is the finding worth sitting with. When agents did misapply their own rules, they went risk-averse 99.5% of the time — suggesting something closer to learned hedging than a random capability failure. Instructing the agent to apply rules mechanically cut misapplication rates roughly in half, which is either reassuring or alarming depending on how you feel about the fact that it needed prompting at all.\n\nThe research is a useful corrective to the common assumption that unfaithful AI behavior has a single upstream cause. Conflating measurement noise with model behavior — especially when using free-text conclusion extraction, which inflated apparent inconsistency to 22-26% — makes it easy to misdiagnose what is actually going wrong.","[\"ai\",\"llm\",\"agents\",\"research\"]","2026-07-08T04:00:00.000Z","2026-07-08T08:43:55.559Z","2026-07-08T08:43:58.385Z","published",null,[],"ai",[24,26,27,28],"llm","agents","research",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.00476",0,{"sections":35},[36,40,45,50,55,60,65,70,75,80,85,89,94,99],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":44},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":46,"slug":47,"count":48,"latest_published_at":49},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":51,"slug":52,"count":53,"latest_published_at":54},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":56,"slug":57,"count":58,"latest_published_at":59},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":61,"slug":62,"count":63,"latest_published_at":64},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":86,"slug":87,"count":83,"latest_published_at":88},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]