{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:J2QAIDUV7JSXVG22VIX4DVT76T","short_pith_number":"pith:J2QAIDUV","schema_version":"1.0","canonical_sha256":"4ea0040e95fa657a9b5aaa2fc1d67ff4c6b0255c65fad3a8c048cc30ad808bbf","source":{"kind":"arxiv","id":"2605.00528","version":2},"attestation_state":"computed","paper":{"title":"SAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU Clusters","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Scheduling AI agent workflows as single units reduces task completion time by 1.64x on GPU clusters","cross_cats":["cs.AI","cs.LG","cs.OS"],"primary_cat":"cs.DC","authors_text":"Dongxin Guo, Jikun Wu, Siu Ming Yiu","submitted_at":"2026-05-01T09:05:28Z","abstract_excerpt":"AI agents execute tens to hundreds of chained LLM calls per task, yet GPU schedulers treat each call as independent, discarding gigabytes of intermediate state between steps and inflating end-to-end latency by 3-8x. We argue that this request-level abstraction is fundamentally mismatched to compound AI workloads, and propose a shift to program-level scheduling: treating the entire agent workflow (not individual inference calls) as the first-class schedulable unit. We present SAGA, a distributed scheduler that implements this abstraction through three mechanisms: (1) Agent Execution Graphs that"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.00528","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.DC","submitted_at":"2026-05-01T09:05:28Z","cross_cats_sorted":["cs.AI","cs.LG","cs.OS"],"title_canon_sha256":"29c1169a251f1f9b1d2f4ce3ac03649d8a0ad6b41c66274100ccb236ea15aa1b","abstract_canon_sha256":"4ee36fde351ce84edcf098394946a2a402b13395be3b520a29cca5a99fa7ae6b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T01:12:07.813739Z","signature_b64":"pg5T0ojyl3JpZN9XM4/2S3Z86+MNsf2Qld51c7A8WBP+hre9QqHQAL18MHFsesGYmhIRqTsNxxp4FTX9E9wwAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4ea0040e95fa657a9b5aaa2fc1d67ff4c6b0255c65fad3a8c048cc30ad808bbf","last_reissued_at":"2026-06-23T01:12:07.813331Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T01:12:07.813331Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU Clusters","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Scheduling AI agent workflows as single units reduces task completion time by 1.64x on GPU clusters","cross_cats":["cs.AI","cs.LG","cs.OS"],"primary_cat":"cs.DC","authors_text":"Dongxin Guo, Jikun Wu, Siu Ming Yiu","submitted_at":"2026-05-01T09:05:28Z","abstract_excerpt":"AI agents execute tens to hundreds of chained LLM calls per task, yet GPU schedulers treat each call as independent, discarding gigabytes of intermediate state between steps and inflating end-to-end latency by 3-8x. We argue that this request-level abstraction is fundamentally mismatched to compound AI workloads, and propose a shift to program-level scheduling: treating the entire agent workflow (not individual inference calls) as the first-class schedulable unit. We present SAGA, a distributed scheduler that implements this abstraction through three mechanisms: (1) Agent Execution Graphs that"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On a 64-GPU cluster serving SWE-bench coding agents and WebArena browser tasks, SAGA reduces task completion time by 1.64x (geometric mean, p < 0.001) over vLLM v0.15.1 with prefix caching and affinity routing, while improving GPU memory utilization by 1.22x and achieving 99.2% SLO attainment under multi-tenant interference.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That real-world agent workflows possess enough predictable structure for Agent Execution Graphs to accurately forecast KV cache reuse across tool-call boundaries, and that the dominant use case is latency-sensitive interactive work where a 30% throughput reduction is acceptable.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SAGA reduces AI agent task completion time by 1.64x on 64-GPU clusters by scheduling at the full workflow level with execution graphs, affinity batching, and completion-time fairness.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Scheduling AI agent workflows as single units reduces task completion time by 1.64x on GPU clusters","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ec3b445e634b92610685497691f42e4ec13f6864cb746bc338fbc38482bf8857"},"source":{"id":"2605.00528","kind":"arxiv","version":2},"verdict":{"id":"7bdc6a0f-cc22-40d5-9f31-deb5a5f8e614","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T19:10:49.517777Z","strongest_claim":"On a 64-GPU cluster serving SWE-bench coding agents and WebArena browser tasks, SAGA reduces task completion time by 1.64x (geometric mean, p < 0.001) over vLLM v0.15.1 with prefix caching and affinity routing, while improving GPU memory utilization by 1.22x and achieving 99.2% SLO attainment under multi-tenant interference.","one_line_summary":"SAGA reduces AI agent task completion time by 1.64x on 64-GPU clusters by scheduling at the full workflow level with execution graphs, affinity batching, and completion-time fairness.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That real-world agent workflows possess enough predictable structure for Agent Execution Graphs to accurately forecast KV cache reuse across tool-call boundaries, and that the dominant use case is latency-sensitive interactive work where a 30% throughput reduction is acceptable.","pith_extraction_headline":"Scheduling AI agent workflows as single units reduces task completion time by 1.64x on GPU clusters"},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.00528/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T19:40:25.577467Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T18:04:38.476505Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"a0717447cd2f49e6112c5df2c7b0134046db947fb0bd82dde60e067808b4e2ab"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.00528","created_at":"2026-06-23T01:12:07.813388+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.00528v2","created_at":"2026-06-23T01:12:07.813388+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.00528","created_at":"2026-06-23T01:12:07.813388+00:00"},{"alias_kind":"pith_short_12","alias_value":"J2QAIDUV7JSX","created_at":"2026-06-23T01:12:07.813388+00:00"},{"alias_kind":"pith_short_16","alias_value":"J2QAIDUV7JSXVG22","created_at":"2026-06-23T01:12:07.813388+00:00"},{"alias_kind":"pith_short_8","alias_value":"J2QAIDUV","created_at":"2026-06-23T01:12:07.813388+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T","json":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T.json","graph_json":"https://pith.science/api/pith-number/J2QAIDUV7JSXVG22VIX4DVT76T/graph.json","events_json":"https://pith.science/api/pith-number/J2QAIDUV7JSXVG22VIX4DVT76T/events.json","paper":"https://pith.science/paper/J2QAIDUV"},"agent_actions":{"view_html":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T","download_json":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T.json","view_paper":"https://pith.science/paper/J2QAIDUV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.00528&json=true","fetch_graph":"https://pith.science/api/pith-number/J2QAIDUV7JSXVG22VIX4DVT76T/graph.json","fetch_events":"https://pith.science/api/pith-number/J2QAIDUV7JSXVG22VIX4DVT76T/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T/action/storage_attestation","attest_author":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T/action/author_attestation","sign_citation":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T/action/citation_signature","submit_replication":"https://pith.science/pith/J2QAIDUV7JSXVG22VIX4DVT76T/action/replication_record"}},"created_at":"2026-06-23T01:12:07.813388+00:00","updated_at":"2026-06-23T01:12:07.813388+00:00"}