Do Not Restart: Residual Completion for Stateful Agent Handoffs
Runzhi Deng, Yiming Zhong, Fang Zhao, Pan Zhou
cs.AI
Quick verdict
CFRC freezes handoff contracts to match strong-agent accuracy at 22–35% cost on five benchmarks; solid mechanism, but headline gaps lack significance tests.
Summary
The paper formalizes stateful agent handoffs—where a cheaper model transfers control mid-task to a stronger one—as commitment-constrained residual completion, and proposes CFRC, a three-stage pipeline. Reached-State Contract Construction (RSCC) freezes, before the successor proposes anything, a residual contract recording accepted bindings, receipt-backed effects, and open obligations. Executable Residual Completion (ERC) has the successor author its entire remaining plan, optionally on an isolated replica, and closes it into a dependency-linked graph. Checked Execution (CE) admits the whole graph against the frozen contract before any live write, rebinds dynamic arguments to native receipts, and accepts success only when live evidence discharges every obligation. The authors prove contract-relative partial correctness, which lifts to the original request only under complete contract construction. Across five stateful benchmarks and five model pairs, CFRC reaches macro accuracy within 1.2 points of strong full-task anchors at 22.0–34.6% of their mean per-surface cost (e.g., 70.8% vs 69.6% at 34.6% cost for GPT-5.4-mini/GPT-5.5). An ablation shows checked whole-graph execution adds 8.8 points over contract-only incremental checking. Audits recover 62/68 ground-truth write obligations and find no ungrounded writes among 50 admitted proposals. Limitations include absent significance testing on headline gaps, an empirically violated completeness assumption (91.2% obligation recall, two false completions), subset-only evaluation of three routing baselines, and a single-tenure cheap-to-strong scope.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CFRC matches strong full-task agents within 1.2 macro points at 22.0–34.6% of their mean per-surface cost across five stateful benchmarks and five model pairs.
- Freezing a residual contract before the successor proposes prevents omission-driven failures that call-level guards cannot detect.
- Whole-graph admission before live writes (CE) adds 8.8 points and saves $8.26 versus step-by-step checking under the same frozen contract.
- Contract-relative partial correctness (Proposition 1) extends to the original residual request under the complete-construction assumption on RSCC.
- Audits find zero ungrounded live writes among 50 admitted proposals (95% Clopper–Pearson upper bound 5.8%), and all 124 CE interventions precede the first live write.
Strengths
Clean problem formalization of an under-served failure mode
Sec. 3.1 defines commitment-constrained residual completion (Eq. 1) separating fixed progress (bindings Bt, receipt-backed effects Kt) from open obligations Ot, with the invoice example making duplicate-payment, invoice-swap, and false-success failures concrete.
Mechanism design that closes a real logical loophole
RSCC freezes the contract before the successor proposes (Sec. 3.2), so an omission cannot hide; Table 3 shows this is cheap ($0.21 mean RSCC spend, $0.00 on three deterministic-schema surfaces).
Genuine matched-frontier ablation
Unusually honest failure reporting with quantified audits
Consistent cost–quality results across model families
Weaknesses
No statistical support for headline gaps
Formal guarantee rests on an assumption its own audit violates
Subset-only baselines weaken comparisons
The handoff trigger is unspecified despite driving the cost story
Single-tenure, simulated-environment scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets stateful tool-agent handoffs, where a cheap model transfers control mid-task to a stronger one and must not restart (duplicate irreversible effects), must not reinterpret user-confirmed choices, and must not claim success while obligations (receipt delivery in the running invoice example) remain open. It formalizes this as commitment-constrained residual completion (Eq. 1, Sec. 3.1) and claims three things: (1) CFRC's three stages — RSCC, ERC, CE — deliver macro accuracy comparable to strong full-task agents at 22.0–34.6% of their cost across five benchmarks; (2) a contract-relative partial-correctness result (Proposition 1) that lifts to the original request under complete contract construction; (3) audits showing the mechanisms work as designed. Scope is explicitly single-tenure, cheap-to-strong, with frozen tasks and public tool schemas only.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CFRC in LLM agent handoffs?
CFRC (Commitment-Frontier Residual Completion) is a three-stage pipeline for handing a task from a cheaper LLM agent to a stronger one mid-task. It first freezes a residual contract listing what is already decided, which effects already happened, and what work remains; the successor then writes its entire remaining plan as a dependency graph; finally a checker admits the whole graph before any real tool call and only declares success when native receipts confirm every remaining obligation.
How much cost does CFRC save compared to just using a strong model?
Across five stateful tool-use benchmarks, CFRC reports macro accuracy within 1.2 points of strong full-task agents while spending 22.0–34.6% of those agents' mean per-surface cost, depending on the model pair (for example, 70.8% vs 69.6% accuracy at 34.6% cost for the GPT-5.4-mini/GPT-5.5 pair).
Does CFRC guarantee the task is completed correctly?
Only conditionally. The paper proves correctness relative to the frozen contract, and this lifts to the original request only if contract construction captured every open obligation. Its own audit recovered 62 of 68 ground-truth write obligations (91.2% recall) and observed 2 false completions out of 50 episodes, so the guarantee is not absolute in practice.
Which benchmarks and models were used to evaluate CFRC?
Five stateful tool-use environments: STATE-Bench, τ²-Bench Retail and Airline domains, ToolSandbox, and Agent-Diff, with model pairs including GPT-5.4-mini/GPT-5.5, GPT-5.6 Luna/Sol, and cross-provider pairs with Claude Sonnet 5, Gemini 3.7 Flash, and Claude Haiku 4.5.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Proposition 1 is stated in Sec. 3.5 but its proof does not appear in the provided text; any proof or appendix was not part of the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.