{"schema_version":"2026-09-05.topic-graph-v1","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu","topic":{"slug":"agentic-cpu","name":"CPU Usage in Agentic AI: Execution, Waiting, and Bottlenecks","description":"What runs outside the model, why a slow agent is not always a busy CPU, and how to read agent-performance claims.","coverage_notes":"A starting collection: one operator architecture account, one experimental preprint, two technical documentation sources, and a vendor opinion. No production-wide CPU-demand estimate, independent benchmark replication, or universal hardware ratio is established. Research gathered on 2026-09-26; no human review is implied.","primer":"Start with where the work runs, then distinguish time spent computing from time spent waiting. The research below keeps experimental results tied to their tasks and treats commercial evaluation proposals as proposals. It is a guide to reading the evidence, not a hardware purchase recommendation.","primer_note":"Manufacturing.ai's own editorial synthesis, not evidence. Every factual statement in it is separately asserted and sourced in this response's concepts and events.","reviewed_through_date":null,"source_count":5,"inspected_source_count":5,"consulted_reference_count":0,"blocked_reference_count":0,"set_aside_reference_count":0},"figures":[],"blocked_references":[],"set_aside_references":[],"concepts":[{"slug":"agent-sandbox-cpu-limits","name":"Check the sandbox's CPU limits","category":"other","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu#agent-sandbox-cpu-limits","assertions":[{"id":"8ca0d9a4-0f40-48cb-ab15-48198e4444fe","predicate":"concept.description","statement":"Linux cgroup documentation distinguishes CPU usage accounting from a configured CPU bandwidth limit. cpu.stat exposes usage and, when applicable, throttling statistics; cpu.max specifies a group's allowed CPU time per period. The documented throttling counters have scheduler and hierarchy limitations. Read both the resource allocation and its accounting before interpreting a container's CPU behavior.","assessment":"reported","editor_reviewed":false,"reference_only_source_count":0,"evidence":[{"id":"e7f1d09c-515e-4bf4-abf0-fc59dbbd1a83","stance":"supports","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Control Group v2: CPU interface files","publisher":"Linux kernel documentation","url":"https://docs.kernel.org/admin-guide/cgroup-v2.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":null,"documents":[]}},{"id":"c40f9e68-187d-4dea-8d49-43833a71d2d0","stance":"qualifies","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Control Group v2: CPU interface files","publisher":"Linux kernel documentation","url":"https://docs.kernel.org/admin-guide/cgroup-v2.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":null,"documents":[]}}]}]},{"slug":"agent-cpu-time-vs-elapsed","name":"CPU time is not elapsed time","category":"other","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu#agent-cpu-time-vs-elapsed","assertions":[{"id":"74640b66-267c-46c0-bc33-9be84b56db10","predicate":"concept.description","statement":"Python's process_time measures the current process's user and system CPU time and excludes sleep. perf_counter measures elapsed duration, including sleep. A slow tool call therefore does not by itself establish high CPU use. A timer around the harness also does not account for CPU work performed by a separate sandbox or remote service.","assessment":"reported","editor_reviewed":false,"reference_only_source_count":0,"evidence":[{"id":"685e21dc-67d5-42cb-a3f4-2505ca645fe7","stance":"supports","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Python time documentation: perf_counter and process_time","publisher":"Python Software Foundation","url":"https://docs.python.org/3/library/time.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":null,"documents":[]}},{"id":"748b898a-2e3d-404a-b5b9-916059bd7b59","stance":"qualifies","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Python time documentation: perf_counter and process_time","publisher":"Python Software Foundation","url":"https://docs.python.org/3/library/time.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":null,"documents":[]}}]}]},{"slug":"agent-completed-work-capacity","name":"Measure completed work","category":"other","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu#agent-completed-work-capacity","assertions":[{"id":"f657c25c-3334-4854-a210-3141395d18c9","predicate":"concept.description","statement":"Intel proposes evaluating completed agent workflows under specified latency, quality, cost, and power constraints, and calls for disclosure of concurrency. This is a vendor's proposed evaluation framework. It does not establish a standard CPU-to-GPU ratio, processor shipment growth, or a comparable result across different agent tasks.","assessment":"reported","editor_reviewed":false,"reference_only_source_count":0,"evidence":[{"id":"2653fb5c-0d2e-40e7-8616-0db20987c000","stance":"supports","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Count the Work Done. Not Just the Cores.","publisher":"Intel","url":"https://www.intel.com/content/www/us/en/newsroom/opinion/count-the-work-not-just-the-cores.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-08-20","documents":[]}},{"id":"a5a3e6fa-eae1-42f4-885b-c0b599eadaa0","stance":"qualifies","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Count the Work Done. Not Just the Cores.","publisher":"Intel","url":"https://www.intel.com/content/www/us/en/newsroom/opinion/count-the-work-not-just-the-cores.html","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-08-20","documents":[]}}]}]},{"slug":"agent-workload-bottlenecks","name":"The bottleneck depends on the task","category":"other","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu#agent-workload-bottlenecks","assertions":[{"id":"dc921d53-18df-4819-bda2-8a8d1d07e8a9","predicate":"concept.description","statement":"In version 3 of Raj and coauthors' preprint, heavy-molecule RDKit tool execution accounts for 85% and 88% of elapsed latency on the two tested systems. Toolformer instead spends about 88% and 77% on model inference. The systems pair Intel Xeon with RTX Pro, and NVIDIA Grace with H200. These workload-specific results do not establish a universal CPU bottleneck or a CPU utilization percentage.","assessment":"reported","editor_reviewed":false,"reference_only_source_count":0,"evidence":[{"id":"b26659b4-d989-4e6e-b486-eb2be9cf0cc8","stance":"supports","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective (v3)","publisher":"Ritik Raj and coauthors, Georgia Tech and Intel / arXiv","url":"https://arxiv.org/html/2511.00739v3","source_type":"academic","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-04-16","documents":[]}},{"id":"0c4f04b2-670d-434e-b8f7-c254dda26071","stance":"qualifies","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective (v3)","publisher":"Ritik Raj and coauthors, Georgia Tech and Intel / arXiv","url":"https://arxiv.org/html/2511.00739v3","source_type":"academic","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-04-16","documents":[]}}]}]},{"slug":"agent-runtime-and-tools","name":"Where the work runs","category":"other","canonical_url":"https://www.manufacturing.ai/topics/agentic-cpu#agent-runtime-and-tools","assertions":[{"id":"30776e8e-6d5c-4ee2-b5ee-7a21ec984cda","predicate":"concept.description","statement":"Anthropic describes Managed Agents as separate session, harness, and sandbox components. The harness calls the model and routes tool requests; the sandbox executes code and edits files. This is an operator's architecture description, not a measurement of CPU consumption. It explains why model serving and tool execution should be examined separately.","assessment":"reported","editor_reviewed":false,"reference_only_source_count":0,"evidence":[{"id":"1f707fb5-cd90-4764-a4ec-3ae12c6bdb5b","stance":"supports","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Scaling Managed Agents: Decoupling the brain from the hands","publisher":"Anthropic","url":"https://www.anthropic.com/engineering/managed-agents","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-04-08","documents":[]}},{"id":"06dba13d-c15b-4cc1-b709-f45920661756","stance":"qualifies","is_primary_for_assertion":false,"origin_independence":"unknown","note":null,"quote":null,"source":{"title":"Scaling Managed Agents: Decoupling the brain from the hands","publisher":"Anthropic","url":"https://www.anthropic.com/engineering/managed-agents","source_type":"other","retrieval_status":"fetched","content_inspected":true,"published_at":"2026-04-08","documents":[]}}]}]}],"events":[],"policy":{"publication_model":"transparency","evidence_entries_may_share_a_source":true,"linked_documents_do_not_add_verification":true,"correction_intake_available":false,"assessment_meanings":{"primary-source confirmed":"Editor-reviewed against a first-party source.","cross-source corroborated":"Editor-confirmed across two independent origins.","reported":"Supported by the attached sources; not yet editor-reviewed.","disputed":"Sources conflict; competing accounts are both attached.","superseded":"Replaced by a newer assertion.","unverified":"No supporting evidence attached."},"evidence_eligibility":"`evidence[]` contains only sources whose content was actually retrieved and read (`content_inspected: true`). Sources that could not be retrieved are counted per-assertion in `reference_only_source_count`, listed for the whole collection in `consulted_references`, and never presented as evidence. `source_count` counts only sources an assertion actually rests on — it deliberately excludes `consulted_references`, so it is never inflated by material nobody could read.","topic_scope":"Assertions and source counts are limited to the topic's current packet lineage. Reusing a concept in another topic does not silently import that other topic's assertions.","rights":"Statements are Manufacturing.ai's own prose. Source excerpts are not redistributed; `quote` is present only where explicitly cleared.","attribution":"Cite as Manufacturing.ai, https://www.manufacturing.ai/topics/agentic-cpu","correction_url":"https://www.manufacturing.ai/methodology#corrections"}}