Do Not Restart: Residual Completion for Stateful Agent Handoffs

1 INTRODUCTION

• At a commitment frontier, accepted choices and realized effects are fixed while request-scoped obligations remain open, so a valid handoff must preserve progress and finish the remainder.In the invoice example, paying INV-42 is already complete, but receipt delivery remains pending; repayment risks a duplicate charge, while declaring success would be premature.Source: §1 · p. 1
• Routing, transcript transfer, planning, and action-level guards address handoff components but do not jointly specify the residual target, ensure complete proposal coverage, and verify realization.A successor can omit an obligation without its proposal revealing the omission, and local guards cannot require an action that was never proposed.Source: §1 · pp. 1–2
• CFRC freezes a residual contract from accepted progress, closes the successor’s continuation into an evidence-linked graph before native authority, and requires live receipts to discharge obligations.The paper establishes conditional partial correctness: realized prefixes preserve encoded progress and returned suffixes satisfy the contract.Source: §1 · pp. 2–3
• 70.8% macro accuracy versus 69.6% for GPT-5.5 at 34.6% mean cost, and 70.8% versus 70.0% with Luna/Sol at 22.0% mean cost.Cross-provider Luna/Sonnet 5 and Luna/Gemini 3.7 Flash remain within 0.5 points of their strong anchors at 26.0% and 44.1% cost; an independent audit recovers 62/68 persistent write obligations and finds no ungrounded writes among 50 admitted proposals.Source: §1 · p. 3

Model routing & agent handoff

• Routing and handoff methods select models or transfer context, but do not specify which effects are fixed or obligations remain open; CFRC constructs that target before proposal.The related-work discussion cites routing and cascades (Ong et al., 2025; Bouchard, 2026; Son et al., 2026; Wei et al., 2026; Liu et al., 2026a) and handoff frameworks (KC & Budathoki, 2026).Source: §2 · p. 3

Classical specifications & recovery

• Specifications, sagas, and plan repair separate execution from requirements, while CFRC builds the residual target from accepted model-generated progress and constrains the successor’s entire proposal.Accepted effects remain preserved, while a failed suffix may return ⊥ without compensation.Source: §2 · p. 3

Stateful control & verification

• Runtime guards constrain proposed actions, but an incomplete prefix cannot reveal future omissions; CFRC checks planned coverage before live writes and requires live evidence for success.Explicit terminal obligations can still detect false completion.Source: §2 · p. 3

3.1 STATEFUL HANDOFF AS A RESIDUAL-COMPLETION PROBLEM

• CFRC treats a handoff as residual completion: fixed progress must be preserved while open request-scoped work is completed.It handles one handoff tenure at a time, does not optimize routing or choose successors, and preserves supported commitments rather than reassessing their desirability.Source: §3.1 · pp. 3–4

3.2 REACHED-STATE CONTRACT CONSTRUCTION

• RSCC grounds accepted history, closes the remaining request into a contract, and freezes it before the successor proposes a continuation.It uses the request, trajectory, reached state, trusted receipts, and public interfaces; freezing prevents proposal omissions but cannot recover obligations missed during contract construction.Source: §3.2 · pp. 4–5

3.3 EXECUTABLE RESIDUAL COMPLETION

• ERC closes the successor’s complete remainder as an evidence-linked graph, exposing coverage and dataflow before any new live write.Replica outputs may support authoring but do not prove live effects; the bridge records supported structure without repairing or replacing task actions.Source: §3.3 · pp. 5–6

3.4 CHECKED EXECUTION

• CE admits the whole graph against the frozen contract before live writes, then realizes calls using live receipts and verifies obligation discharge.Completion requires matching native evidence; mutable-state obligations must remain valid, and a failure after a state-changing call ends the tenure without implying rollback.Source: §3.4 · pp. 6–7

3.5 CONDITIONAL END-TO-END GUARANTEE

• Under complete RSCC construction and the stated ERC and CE assumptions, a returned suffix preserves frozen progress and completes every in-scope residual obligation.Without complete construction, the guarantee is contract-relative; it does not certify the constructor, guarantee a valid proposal will be found, or imply rollback when ⊥ is returned.Source: §3.5 · p. 7

4 EXPERIMENTS

• The five-benchmark evaluation uses frozen tasks, public tools, and fixed policies, with two primary model pairs and three cross-provider pairs.CFRC accesses public action schemas, effect types, and native receipts, without simulator internals, reference outputs, or evaluator rewards.Source: §4 · p. 7

4.1 END-TO-END QUALITY AND COST

• CFRC matches strong anchors within 1.2 points at 34.6% mean cost for GPT-5.4-mini/GPT-5.5 and within 0.9 points at 22.0% for Luna/Sol.Table 1 reports end-to-end results across five surfaces; macro standard deviations across five runs are 0.54 and 0.58 points, respectively.Source: §4.1 · pp. 7–8, Table 1 · p. 8
• On StepWise, AgServe-QAC, and MTRouter subsets, CFRC improves accuracy by 3.4, 7.1, and 6.3 points while cutting cost by 37–49%.RouteLLM-BERT and FrugalGPT trail GPT-5.5 by 2.0 and 1.3 points at 59.7% and 96.3% cost, while CFRC gains 1.2 points at 34.6% cost.Source: §4.1 · p. 8
• Luna/Sonnet 5 and Luna/Gemini 3.7 Flash match strong anchors within 0.5 points at 26.0% and 44.1% cost, while Haiku 4.5/Sol trails Sol by 5.0 points at 57.4% cost.CFRC gains 17.0 points over the standalone Haiku 4.5 anchor; ToolSandbox similarity softens under Gemini.Source: §4.1 · p. 8

4.2 VALIDATING THE THREE STAGES

• Matched-frontier accuracy rises from 58.9% for Direct Takeover to 62.1% for Certified Direct, 65.4% for Contract Plan, and 70.8% for Full CFRC.Full CFRC costs $38.80; compared with Certified Direct, whole-graph admission adds 8.8 points and saves $8.26 under the same contract.Source: §4.2 · pp. 8–9, Table 2 · p. 8
• RSCC averages $0.21 across surfaces, and selective invocation engages on 59.6% of episodes as three call mechanisms drive 43.6% savings over Direct Takeover.RSCC eliminates exploratory re-querying with grounded obligations, while fail-closed CE checks halt ungrounded plans early to prevent runaway retries.Source: §4.2 · p. 9, Table 3 · p. 9

4.3 AUDITING CONTRACTS AND CHECKED EXECUTION

• Contract audits recover 62 of 68 persistent write obligations (91.2% recall), and admit no ungrounded live writes among 50 audited proposals.Six omissions involved implicit updates without entity mentions; four were blocked downstream and two yielded false completions. Formal lift to the residual objective remains conditional on complete construction because conversational disclosures fall outside tool schemas.Source: §4.3 · p. 9, Table 4 · p. 9
• All 124 Checked Execution interventions occur before the first live write, including 43 proposal rejections and 81 holds.The 43 rejections cover 25 coverage, 10 dependency, and 8 grounding failures, and all reproduce under deterministic replay.Source: §4.3 · p. 9, Table 4 · p. 9

• CFRC formalizes stateful handoff as commitment-constrained residual completion, using frozen residual contracts, closed graphs, and live receipts to match strong full-task quality at a fraction of inference cost while blocking unadmitted mutations.Source: §5 · p. 9
• Evaluation covers single-tenure cheap-to-strong handoffs under frozen contracts; recursive routing is outside immediate scope, and formal lift depends on complete RSCC construction.Source: §5 · p. 9
• The authors state that code and schemas will be open-sourced.Source: §5 · p. 9