Agent-Integrated Software: Interaction Contracts and Continuous Assurance
Shengcheng Yu, Chunrong Fang, Zhenyu Chen
cs.SE · cs.AI
Quick verdict
Perspective formalizing agent-in-app delegation as interaction contracts plus versioned assurance; careful and honest, but unimplemented and unvalidated.
Summary
This position paper addresses the coordination problem that arises when a built-in LLM agent shares application state with users who keep interacting through a conventional GUI: delegated work continues while its meaning changes. It defines Agent-Integrated Software (AIS), a pattern combining a conventional core, direct interaction, and a built-in agent, and Intent-Level Interaction Abstraction (IIA), a task-level transition system over goals, revisions, bindings, endorsements, and outcomes. AIS execution is modeled as an open labeled transition system; conformance is stated as inclusion of projected concrete traces within traces permitted by an interaction contract ⟨Pre, Step, Inv, Post, Dep⟩. A compact reviewed-disclosure contract with four clauses (endorsement binding, current conjunctive authority, gate-controlled admission, outcome fidelity) instantiates a running meeting-materials scenario, and a gate-induction proposition shows admission safety is independent of the planner. Continuous assurance is framed as versioned claim records (supported/refuted/insufficient) with a dependency-based reassessment rule. The paper derives eight engineering implications, four design principles, six research directions, and four hypothetical domain probes. Limitations: no implementation, runtime, or empirical evaluation; the formalism is standard refinement and contract theory with near-immediate propositions; all machinery is instantiated on a single scenario; and recovering the required abstractions from real applications is deferred to future work.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
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
- Dependable agent integration requires an explicit, maintained correspondence (interaction–effect obligations, Eq 4) between revisable task intent and application behavior; local component validity — valid API calls and authenticated permissions — does not establish it (Prop 2.1).
- AIS (conventional core + direct interaction + built-in agent) and IIA can be formalized as an open transition system whose projected traces must be contained in the traces permitted by an interaction contract ⟨Pre, Step, Inv, Post, Dep⟩ (Eq 5).
- Selected admission invariants (K1–K3) can be enforced atomically by a trusted host gate independently of how the LLM planner selected the action (Prop 4.1), separating planning nondeterminism from admission safety.
- Continuous assurance should be maintained as versioned, scoped claim records with supported/refuted/insufficient statuses and dependency-triggered selective reassessment (Eq 8–9, Table 4), rather than generic task-success metrics.
Strengths
Well-motivated, precisely isolated problem
Sec 1's meeting-materials scenario — a GUI edit removes recipient b after an endorsement of revision 7, and a delivery timeout leaves disclosure unknown — crisply exhibits the gap between valid API calls plus authenticated permissions and interaction–effect conformance, formalized as the Prop 2.1 counterexample (Sec 2.3).
Disciplined conceptual hygiene
The paper consistently separates admission from occurrence (Sec 2.1), endorsement from agent inference (Sec 3.5), safety conformance from progress and goal adequacy (Sec 2.3, 6.3), and evidence states supported/refuted/insufficient (Eq 8, Sec 4.4) — the trichotomy explicitly prevents missing evidence from being read as a passed check.
Actionable gate-separation result
Coherent synthesis structure
Honest scope statement
Weaknesses
No implementation or empirical evidence
Thin formalism with near-trivial results
Single-scenario instantiation
The hardest problem is deferred
Heavy self-citation cluster
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position/perspective paper (arXiv 2609.11381, cs.SE) arguing that embedding an LLM agent into an existing application creates a persistent coordination problem — users revise goals and manipulate shared objects while delegated execution continues — that cannot be assessed by response quality alone. The authors (TUM; Nanjing) propose **Agent-Integrated Software (AIS)**, a pattern of conventional core + direct interaction + built-in agent, and **Intent-Level Interaction Abstraction (IIA)**, a task-level transition system over goals, revisions, bindings, endorsements, and outcomes. Central claims: (1) local component validity does not establish interaction conformance (Prop 2.1, via the stale-endorsement counterexample of Sec 2.3); (2) conformance is finite-trace inclusion of projected concrete traces within contract-permitted traces (Eq 5); (3) admission invariants K1–K3 can be enforced by a trusted host gate independently of the planner (Prop 4.1); (4) assurance claims should be versioned records with supported/refuted/insufficient statuses and dependency-triggered reassessment (Eq 8–9, Table 4). Scope is explicitly conceptual: Sec 6.4 disclaims implementation, empirical results, and cross-domain validation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 Agent-Integrated Software (AIS)?
AIS is a software pattern proposed in this arXiv position paper (2609.11381): an application that combines a conventional core (owning domain objects, business rules, and durable state), a direct interaction path (the normal GUI), and a built-in LLM-based agent that plans and executes delegated goals. The paper stresses that architectural membership alone guarantees nothing — a defective integration is still an AIS instance; dependability must come from separately specified interaction contracts.
What is Intent-Level Interaction Abstraction (IIA)?
IIA is the paper's task-level abstraction of delegated work: an abstract transition system whose states summarize goals, revisions, contextual bindings, delegated scope, and unresolved decisions, and whose events cover goal submission, revision, endorsement, intervention, admission, and outcome reporting. It defines what users can inspect and control about a running agent task, independent of whether the surface is a chat panel, an editable preview, or existing GUI controls.
Does the paper include an implementation or experiments?
No. The authors state explicitly (Section 6.4) that the work provides definitions, conditional propositions, worked examples, and a research agenda — no implemented runtime, no empirical results, and no cross-domain validation. Comparative evaluation with and without the contract artifacts is proposed as future work.
What is 'continuous assurance' in this paper?
A scheme for maintaining claims about an agent-integrated application as versioned records A = ⟨q, Ω, D, E, R, S⟩ where the status S is supported, refuted, or insufficient. When a dependency changes (model swap, policy change, adapter update), a rule triggers reassessment of only the affected claims (Eq 9). Crucially, missing evidence yields 'insufficient', never a silent pass — the paper's five claims Q1–Q5 (Table 4) cover admission authority, effective stopping, outcome reporting, continuity, and release behavior.
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.
- Figures 1–4 are extracted as interleaved diagram-label text plus captions; the visual layout is lost, but the captions and body prose fully determine the interpretation used in this review. The text otherwise appears complete through the final reference [55].
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.