{"resourceId":"avatar-scientific-workflow-orchestration-20260909","versions":[{"version":"external-ad8ab2150ae571cb5489a4fad0f4793b657db1b62bb097fcc707fcbe1e2ee94c","resource":{"id":"avatar-scientific-workflow-orchestration-20260909","title":"Avatar (UChicago/ANL): LLM workflow orchestration cuts reported compute wastage 55% and GPU-busy time 40%","organization":"University of Chicago and Argonne National Laboratory","sector":"Academic research computing / scientific workflows","geography":"United States (University of Chicago; Argonne National Laboratory)","publishedAt":"September 9, 2026 (arXiv 2609.10509v1)","publicationDate":"2026-09-09","eventDate":"2026-09-09","sourceName":"arXiv","sourceLabel":"Academic preprint — agentic scientific workflow orchestration evaluation","sourceUrl":"https://arxiv.org/html/2609.10509v1","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["developers-agents","infrastructure","operating-model","knowledge-work"],"finding":"Avatar (Raj et al.; UChicago + Argonne; 9 Sep 2026) proposes an actor-based architecture (orchestrator, executor, provenance monitor) with pluggable rule or LLM decision policies on Academy/Parsl/TaskVine. Across three workloads, rule mode reproduces native execution on one unchanged core; LLM-backed Avatar reports a 55% reduction in compute wastage and a 40% cut in GPU-busy time. Emphasizes adapter-validated action catalogs and provenance triggers to bound agentic risk.","sledRelevance":"Directly relevant to university/national-lab research computing seeking agent help for workflow orchestration without abandoning provenance and action boundaries.","evidence":"arXiv HTML: UChicago/ANL authors; three workloads; rule mode ≡ native; LLM-backed −55% compute wastage, −40% GPU-busy time; Academy 0.5.0, Parsl 2026.3.9.","architectureImplications":"Introduce agents as pluggable policies behind validated action catalogs and provenance monitors—not as unconstrained shell access on HPC workflows.","governanceImplications":"Agentic orchestration needs explicit approval of allowed actions and audit of provenance diagnoses before production scheduler control.","securityPrivacyImplications":"LLM-backed orchestrators calling external model APIs must control workflow metadata egress and credential scope on shared HPC.","caveats":"Preprint three-workload author evaluation; not multi-campus production evidence; OpenAI client used for agent reasoning.","streamIds":["research"],"roles":{"sales":"Sales framing: Research computing teams drowning in workflow failures want agent help that still respects HPC controls. Avatar supports a bounded pitch: pluggable LLM orchestration with measured efficiency claims (−55% compute wastage, −40% GPU-busy time on three workloads) while rule mode matches native execution—not proof every campus WMS will see the same gains. Ask which workflows waste GPU time today, what actions agents may take, and how provenance is reviewed. Unsupported: universal 55%/40% savings, safe unconstrained autonomy, or replacement of workflow engineers.","engineering":"Engineering framing: Fit is an actor architecture (orchestrator/executor/provenance) with rule or LLM policies over Academy/Parsl/TaskVine. Integration points are existing WMS interfaces, validated action catalogs, and TaskEvent provenance streams. Prerequisites: instrumented workflows, LLM API governance, and failure-diagnosis policies. Constraints: preprint three-workload results; OpenAI client dependency in the reference implementation. Security must scope credentials and metadata sent to model APIs. Proof of value is wastage/GPU-busy metrics on local pilots versus native baselines.","delivery":"Delivery framing: Pilot Avatar-style bounded orchestration on one workflow family with rule-mode parity tests before enabling LLM policy. Dependencies: research computing, workflow owners, and AI/API governance. Ownership: RC owns runtime; science teams own workflow correctness. Skills adoption trains operators on action-catalog changes and provenance review. Governance checkpoints: allowed-action approval, API egress review, rollback to rule mode. Acceptance: native parity in rule mode plus measured efficiency on the pilot; risks include opaque LLM decisions and API data leakage."},"roleDetails":{"sales":{"customerProblem":"Scientific workflows waste GPU/CPU time under static rules while teams fear unconstrained LLM control of HPC.","stakeholders":["Research computing","Workflow/WMS owners","Lab computational scientists","Security"],"discoveryQuestions":["Which workflows waste the most GPU time?","What actions may an agent take?","How is provenance reviewed?","What is the native baseline?"],"valueHypothesis":"Bounded LLM orchestration (as Avatar reports −55% wastage / −40% GPU-busy on three workloads) can improve efficiency when action catalogs and provenance gates exist—without claiming universal savings.","potentialEngagement":"Bounded orchestration pilot: select one workflow, define action catalog, compare rule vs LLM policy on wastage metrics.","unsupportedClaims":["Guaranteed 55%/40% savings everywhere","Agents can run without action limits","Replaces workflow engineers"]},"engineering":{"fit":"Actor-based WMS overlay with pluggable rule/LLM policies and provenance monitoring on Parsl/Academy-class stacks.","architectureIntegration":"Wrap orchestrator/executor/provenance actors around existing WMS; route agent decisions through validated actions only.","prerequisites":["Instrumented TaskEvent streams","Action catalog","LLM API access policy","Native baseline runs"],"constraints":["Author three-workload evaluation","Reference uses OpenAI client","WMS-specific adapters needed"],"security":"Minimize workflow metadata to external LLMs; scoped credentials; audit agent actions.","proofOfValue":["Rule-mode native parity","Wastage and GPU-busy deltas vs baseline","Provenance completeness checks"]},"delivery":{"work":"Run rule-mode parity, then limited LLM-policy pilot with rollback and metric dashboards.","dependencies":["RC platform team","Science workflow owners","API/security review"],"ownership":"Research computing owns orchestration runtime; PIs own scientific correctness.","skillsAdoption":"Train operators to edit action catalogs and read provenance diagnoses.","governanceCheckpoints":["Allowed-action sign-off","API egress review","Rollback drill"],"acceptanceCriteria":["Native parity in rule mode","Documented efficiency metrics on pilot","Audit log of agent actions"],"risks":["Opaque LLM decisions","Metadata leakage to APIs","Overfitting to demo workloads"]}},"retrievedAt":"2026-09-23T22:09:53-05:00","enrichedAt":"2026-09-23T22:09:53-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Efficiency gains help shared HPC users if pilots include facilitator support; otherwise only teams already using advanced WMSs benefit.","procurementImplications":"Evaluate agentic orchestration by required action-catalog and provenance features plus local baseline metrics, not vendor autonomy slogans.","operatingModelImplications":"Keeps human-defined action boundaries while allowing LLM policy experiments; ops must fund monitoring and rollback.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2609.10509v1","referenceExcerpt":"Avatar’s rule mode reproduces native execution, with a single unchanged core running all three","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}