acceptodds
Under review as a conference paper at ICLR 2027

Do Not Restart: Residual Completion for Stateful Agent Handoffs

Abstract

Routing and cascades reduce tool-agent cost by transferring control across models, but stateful handoffs must preserve accepted choices, realized effects, and unfinished obligations. We formulate this as commitment-constrained residual completion and introduce Commitment-Frontier Residual Completion (CFRC). CFRC enforces target-before-proposal, whole-proposal-before-authority, and live-evidence-before-success: it freezes a residual contract from accepted progress, closes the successor continuation into an evidence-linked graph, and admits execution only when the remainder is covered, with live receipts discharging obligations. We establish contract-relative partial correctness, which extends to the original residual request under complete contract construction. Across five environments and two same-provider model pairs, CFRC achieves comparable macro accuracy to strong full-task agents at 22.0%–34.6% mean per-surface cost, with additional cross-provider results demonstrating broader transfer.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, agent handoffs, stateful tool use, residual completion, tool-use verification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_RZzLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how an agent can take over a partially completed task while preserving completed work and fulfilling remaining requirements. CFRC constructs a residual contract, checks a complete continuation plan, and verifies execution using tool evidence. Experiments report reduced API costs with broadly comparable macro-level task performance.

Critical strengths

-- Relevant research problem. Preserving progress and avoiding omissions during agent handoffs is a practical problem.

-- Useful empirical evidence. The reported cost performance results are promising, and the component ablations provide preliminary support for several design choices.

Critical weaknesses

-- The method is difficult to understand and reproduce. Dense terminology and abstract descriptions makes it very hard to understand the actual workflow. Contract construction rules, prompts, tool adapters, and handoff criteria are not sufficiently specified. Although the invoice illustration explains the motivation, a complete benchmark case tracing the contract, proposed actions, checks, and execution outcome is missing. These gaps make it difficult to assess how the system operates in practice. And there is no code for me to check.

-- The experiments do not fully align with the mechanism claims. The results support the potential of the overall system, but do not cleanly establish why it works. For example, the comparison between Certified Direct and Full CFRC in Table 2 changes both ERC and CE, so the improvement cannot be attributed solely to whole-graph admission. More targeted evidence is needed to substantiate this central claim.

Questions for authors

-- Could you provide a complete benchmark case showing the request, constructed contract, proposed plan, checks, and final outcome?

-- Could you specify the construction rules, prompts, tool adapters, and handoff criteria needed to reproduce the evaluated implementation? Or maybe provide the code/relevant data?

-- What evidence isolates the benefit of whole graph admission from plan construction and the other execution checks?

Reviewer_4WbjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how one model can take over an ongoing agent task while preserving confirmed choices and completed actions and fulfilling remaining requirements. It proposes Commitment-Frontier Residual Completion (CFRC), which constructs a frozen contract from the accepted history, checks the successor’s complete plan before allowing live writes, and verifies completion using actual execution evidence. The method guarantees correctness relative to the contract under stated assumptions; correctness for the original task additionally requires complete contract construction. Across five benchmarks, CFRC achieves comparable average task success to strong-model agents at 22.0%–34.6% of their mean per-benchmark inference cost for the two primary model pairs.

Critical strengths
  1. The paper studies a timely problem of commitment-constrained residual completion, by introducing the target-before-proposal, whole-proposal-before-authority, and live-evidence-before-success pipeline. This cleanly separates specification (RSCC), planning (ERC), and authorization/realization (CE).
  2. Evaluates on five stateful benchmarks, multiple model pairs, and cross-provider transfers. It is also cost-aware and uses token cost as the measure.
  3. The problem setting, pipeline, and running example are clear, with helpful diagrams and a consistent notation for contracts, graphs, and evidence.
  4. It provides a generalizable control pattern that freeze residual target, exposes the full plan and gate effects behind verification, which should influence how multi-model agents are engineered for persistent environments.
Critical weaknesses
  1. Experiment design
  • According to table 6, except the τ2\tau^2-Airline environment, the successor can author its plan on an isolated replica, and observe intermediate results before committing. As far as I can tell, none of the baselines or full-task anchors have this capability. How much of CFRC’s gain over the strong full-task agent (+1.2 / +0.9 macro) comes from replica-based trial-and-error rather than from the residual contract itself?
  • The choice of budgets is not sufficiently explained. The choice of budgets is not sufficiently explained. Budgets vary across environments and model pairs: STATE allows 2 residual calls, while Airline task allows 4 or 5. The paper claims these budgets are frozen, but does not clearly justify how they are selected. I would recommend documenting this selection procedure and reporting result sensitivity to these budgets.
  • The baseline comparisons could be stronger and more consistent. StepWise, AgServe-QAC, and MTRouter each cover only two environments, and external baselines do not consistently receive the same five-run evaluation as CFRC.
  • The results need a breakdown by whether a handoff occurred. For example, in the Retail environment, only 20% of the episodes invoke the strong model, yet the accuracy rises from the cheap agent's 65% to 80%, which even exceed the strong agent's 57%. This is possible if handoffs fix most of the selected failures, but the aggregate results do not show whether that explains the gain. Could report report success rates separately for handed-off and non-handed-off episodes, and clarify whether CFRC’s cheap controller uses the same prompts, tools, and budgets as the cheap full-task baseline?
  1. Method generalizability and human prior
  • The method mainly relies on hand-written environment adapters. Each environment requires separate definitions of effect types, receipt recognizers, and terminal predicates. In three structured environments, obligations are extracted through deterministic rules, meaning no model inference cost. These adapters provide structured domain knowledge that the baselines may not receive in an equivalent form.
  • It is unclear how CFRC handles new information that changes the plan. CFRC checks the full remaining plan before live execution, but some tasks require later decisions based on tool results or further user input. For example, a failed payment may require the user to choose another payment method. Could you clarify how CFRC revises the plan in such cases and evaluate the additional cost and effect on task success?
  • The results depend on a fixed handoff rule. CFRC switches to the strong model once a confirmed choice or completed action leaves further work to do, even if the cheap model could continue. It is unclear whether the same benefits hold when handoffs are triggered by uncertainty or failure. Please test these alternative handoff policies to show how broadly the results apply.
  1. Limited theoretical contribution
  • Could the authors clarify what Proposition 1 establishes beyond the checker’s soundness assumptions? The result appears to follow largely from assuming that the checks and execution are sound, while complete contract construction, which is the harder step, is assumed in A1. What is the additional insight the proposition provide?
  • What is the role and novelty of Lemma 2? It appears to formalize the intuitive point that a checker cannot distinguish cases when the information needed to distinguish them is unavailable.
  1. Computation of cost
  • How sensitive are the cost savings to promotional pricing? The reported costs use promotional rates for Sol and Gemini. While these rates reflect the stated pricing snapshot, would it be better to report token usage by model?
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.