{"paper":{"title":"FlowSteer: Towards Agents Designing Agentic Workflows via Reinforced Progressive Canvas Editing","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A single agent can design complete agentic workflows end-to-end by making sequential edits to an executable canvas that supplies real-time syntax-checked feedback.","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Erik Cambria, Haoran Luo, Mingda Zhang, Qika Lin, Rui Mao, Tiesunlong Shen, Wenjin Liu, Xiaoying Tang","submitted_at":"2026-02-02T05:30:42Z","abstract_excerpt":"In recent years, agentic workflows have been widely applied to solve complex human tasks. However, existing workflow construction still faces key challenges, including human-dependent workflow construction, the lack of graph-level execution feedback, and the inability to repair errors in-loop during long-horizon construction. To address these challenges, we propose FlowSteer, a new paradigm of Agent Designing Agentic Workflows - a single agent itself end-to-end designs the workflow that a downstream executor runs. To support this paradigm, we introduce the Workflow Canvas, a novel executable g"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experimental results on twelve datasets show that FlowSteer significantly outperforms baselines across various tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That real-time syntax-checked execution feedback from the Workflow Canvas is sufficient to train a policy agent that can reliably repair errors during long-horizon workflow construction without external human guidance.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A reinforcement learning policy agent designs executable agentic workflows by issuing atomic edits to a feedback-providing Workflow Canvas environment.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A single agent can design complete agentic workflows end-to-end by making sequential edits to an executable canvas that supplies real-time syntax-checked feedback.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2b56aa3efe74f45787fd0c5caf250bc3e668b37b905735c3a90ac8a8c94ba93c"},"source":{"id":"2602.01664","kind":"arxiv","version":4},"verdict":{"id":"cf8b4ae0-4740-41ec-b9cc-09e4164b647b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T08:45:54.061811Z","strongest_claim":"Experimental results on twelve datasets show that FlowSteer significantly outperforms baselines across various tasks.","one_line_summary":"A reinforcement learning policy agent designs executable agentic workflows by issuing atomic edits to a feedback-providing Workflow Canvas environment.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That real-time syntax-checked execution feedback from the Workflow Canvas is sufficient to train a policy agent that can reliably repair errors during long-horizon workflow construction without external human guidance.","pith_extraction_headline":"A single agent can design complete agentic workflows end-to-end by making sequential edits to an executable canvas that supplies real-time syntax-checked feedback."},"references":{"count":21,"sample":[{"doi":"","year":null,"title":"Greedy always picks largest coin≤N, which is locally optimal but not always globally optimal","work_id":"6358b90a-6834-4909-b15d-fe15e73c2df4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Greedy fails when using fewer large coins plus more medium coins yields fewer total coins (e.g., N=30: greedy=25+5×1=6 coins, optimal=3×10=3 coins)","work_id":"8a4dbd32-6020-4b30-bf94-22ec19bb3fa1","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Dynamic programming guarantees finding the true minimum coin count","work_id":"ca908b1c-1bbd-4c05-abaa-571c89dadfa0","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The coin system {1,10,25} is NOT canonical (unlike {1,5,10,25} US coins), so greedy can fail. Plan:","work_id":"83aaf664-afdc-4282-ab47-6414d4268a43","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Implementgreedy_coins(n)that iteratively subtracts largest possible coin","work_id":"b8e9206a-1666-40e7-8b90-b93558734947","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":21,"snapshot_sha256":"d2b3a863eeefa54eefea25d50c3acddc941a82b3c63456488e6c6641ff07bf5a","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"}