acceptodds
Under review as a conference paper at ICLR 2027

Skill-in-Skill: Agent Skills That Carry Their Own Execution Conditions Across Domains

Abstract

Agent skills provide LLMs with reusable procedural knowledge, but existing skills typically specify what to do without carrying the conditions that govern valid use. As a result, activation scope, authorization, and execution evidence remain external to the skill and may be lost when the skill is transferred or reused. We introduce Skill-in-Skill (SIS), a portable skill architecture that binds a reusable capability with its governance contract as a single self-contained unit. SIS carries activation, authorization, execution, and verification conditions with the skill, and records runtime evidence to support independent validation of governed execution. We compare SIS with an External Governance Sidecar (SIDECAR) that keeps equivalent governance outside the skill package. Across open-weight models, SIS consistently reduces context-access and tool-use overhead, with up to 90.3% fewer repeated task-context reads and 13.2% lower end-to-end latency. In cross-boundary transfer tests, SIS preserves authorized execution after site or principal changes while blocking effects when destination governance is absent. These results show that binding governance to the capability improves the portability and efficiency of governed skill execution without weakening enforcement.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Model Agents, Agent Skills, Skill Governance, Runtime Verification, Authorization and Access Control, Portable Agent Architecture, Cross-Boundary Transfer, Execution Evidence and Auditing, Tool-Use Efficiency, Secure Agent Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2LCcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper argues that agent Skills carry their capability but not their governance. Activation scope, authorization, and execution evidence live in the host environment and can be lost when a Skill is transferred, a problem the authors call the Capability–Governance Gap. They propose Skill-in-Skill (SIS), which packages a capability K\mathbf{K} with execution constraints C=(T,Σ,Ω)\mathcal{C}=(T,\Sigma,\Omega) and governance rules G=(s,α,v)\mathcal{G}=(s,\alpha,v) as W(K)=(K,C,G)\mathcal{W}(\mathbf{K})=(\mathbf{K},\mathcal{C},\mathcal{G}). In SIS, activation requires task match, scope, and authorization; state-changing actions require destination-issued grants; and a six-stage evidence trail supports verification. SIS is compared with an External Governance Sidecar (SIDECAR) that holds the same contract outside the package. The evaluation uses 120 self-constructed business-workflow tasks (10 categories, 3 repeats, three hosted models). On these tasks, SIS reduces per-run file reads by 7.6–11.2%, tool calls by 2.6–9.1%, and cumulative task-context.json reads by up to 90.3%, while verified completion is essentially unchanged. In a transfer test on DeepSeek (six trials per cell), SIS reaches 100% authorized completion after site rebinding and principal delegation, while SIDECAR is blocked in every trial; both are blocked when destination governance is absent.

Critical strengths

S1. Clean controlled comparison. SIS and SIDECAR share capabilities, contracts, tools, permissions, and budgets, so only placement differs. Results are paired per task, use bootstrap intervals, and include held-out instances. Per-class estimates are reported in full, including unfavorable ones such as DeepSeek's increases. This makes the results easy to interpret.

S2. Reasonable problem framing. Distinguishing capability from governance is a sensible way to think about portable Skills. So are separating applicability, scope, and authorization (TT vs. ss vs. α\alpha), separating attempted actions from admitted effects, and returning unknown instead of inferring authorization from task success.

Critical weaknesses

W1. The contribution amounts to relocating the governance specification into the Skill package. The paper states that the sole architectural difference between SIS and its baseline is placement (l. 266–267). The formalism names the fields of the package but establishes no property; Eq. 4 is simply a conjunction of three predicates. Existing work already makes governance explicit through permission manifests and runtime enforcement, e.g., SkillGuard, Progent, Proof-Carrying Agent Actions, and ClosureBound, all of which the paper cites. Relative to these, the only new element is co-location, and none of them is compared empirically.

W2. The efficiency gains are small and appear to be an artifact of how the baseline exposes governance. In absolute terms, per-run file reads fall by 0.28–0.48 against a SIDECAR mean of about 4, and tool calls fall by 0.27–0.93 against a mean of about 10. The headline "up to 90.3%" is a cumulative count for a single file over 360 runs: 93 → 9 reads for MiMo, or about 0.23 reads per run. For DeepSeek, the main backbone, output tokens, cache-read tokens, and latency all increase (+3.2%, +2.6%, +2.6%). Fig. 6 shows where the saving comes from: task-context lookups fall from 198 to 24, while skill-instruction reads stay at 246 → 246. In other words, the governance content is loaded together with the Skill instead of being looked up later. A sidecar that injects the external contract into context once at load time would likely achieve the same saving. The paper neither tests this nor specifies where SIDECAR stores its contract.

W3. The main experiment never measures governance effectiveness. The central claim concerns governance, yet the 2,160-run study reports only resource use and verified completion. Completion is indistinguishable between placements (360/360 vs. 360/360, 355 vs. 356, 354 vs. 355). The tasks are stratified into normal, boundary, conflict, and reject instances (App. A), but no per-stratum results are reported. The paper also reports no counts of violation attempts, no blocked unauthorized actions, and no per-check verifier outcomes (Table 6). The evidence therefore shows only that SIS does not hurt completion, not that it governs execution better.

W4. The only governance experiment is true by construction. The 100% vs. 0% result in Table 5 follows directly from the setup. App. G states that SIDECAR "lacks the self-contained credentials required to satisfy destination gatekeeping, resulting in predictable pre-effect containment." The test also uses a single model and six trials per cell. It includes no adversarial conditions, such as tampered packages, malicious Skills, or prompt injection, even though Sec. 2 cites the relevant security literature.

W5. The design leaves the self-assertion problem unresolved. If the carried G\mathcal{G} determines what a Skill may do, a malicious Skill can declare permissive governance for itself. If instead the destination's local policy decides (l. 226–227), the carried governance is only advisory, and it is unclear what portability adds over destination policy. The paper does not discuss this tension. App. G mentions "cryptographic provenance" but does not specify signing, issuers, or a trust chain.

Questions for authors

Where does SIDECAR store its contract? Is it inside task-context.json or read alongside it? How does SIS compare with a sidecar that is loaded into context once at Skill-load time?

Who signs or certifies the carried governance? How does the destination prevent a Skill from granting itself broader scope than local policy permits? If local policy always prevails, what does carried governance add?

1 comment

Consider the site-rebinding example. A skill prepared for account A arrives at a site serving account B. The receiver needs to identify the capability version, the contract governing its requested action and the verifier for its effects. SIS makes that association part of the transferred skill. B still decides whether this actor may write to B. This distinction explains both the benefit of carrying governance and why a skill cannot authorize itself.

W1/W5. What travels, and who trusts it? Equation 4 states the activation conditions. The operational contribution is what follows: the destination compares the presented capability, contract and verifier digests with its own expected values, checks the actor and requested scope, and issues a local grant. Simply declaring a broader scope in the package cannot change those expected values or produce a grant. When destination governance was withheld, both placements were blocked before effect in all six trials. Appendix G describes signed delegation, but does not specify a general package signer or certification chain. SkillGuard, Progent, Proof-Carrying Agent Actions and ClosureBound motivate the use of manifests, mediated actions and evidence. Our specific contribution is the unit that binds these checks to a transferable agent skill, together with a measurement of what changes when that binding is carried inside the unit.

W2. Where the contract lives and how many accesses it saves. We checked the staged files in a matched pair of formal MiMo trials. SIDECAR stored the contract at stage/sidecar/governance.json with an adjacent binding.json; SIS stored the pair under stage/_sis/. Each task prompt explicitly pointed the agent to its respective binding file. Task-context reads in Table 3 are a separate access measure. On MiMo they fell from 93 to 9 over 360 runs, or 0.233 fewer reads per run for that file. Across models the total saving was 0.278–0.475 file reads and 0.269–0.933 calls per run. Figure 6 localizes the Qwen change particularly clearly: task-context reads fell from 198 to 24 while skill-instruction reads stayed at 246. One-time injection of an external contract could recover the same lookup saving. The study measures the two implemented paths. We also agree that DeepSeek’s increases in output tokens, cache-read tokens and wall time preclude a cross-model claim about those endpoints.

W3. What the 2,160 runs say about governance. Verified completion includes checks on actor identity, accepted-effect authorization and write scope as well as artifact and final-state correctness (Table 6). In SIDECAR/SIS order, verified completions were 360/360 versus 360/360 on DeepSeek, 355/360 versus 356/360 on MiMo, and 354/360 versus 355/360 on Qwen. The supported claim is that the measured access savings coexist with similar observed verified outcomes. Aggregate completion cannot tell us how often agents attempted violations, which verifier checks failed or how those events were distributed across normal, boundary, conflict and reject fixtures. Table 5 supplies direct evidence about attempted writes for the separate transfer experiment.

W4. What the 6/6 versus 0/6 contrast establishes. The external binding stayed at the origin in the tested transfer protocol. A sidecar that travels with the capability and is successfully reassociated could satisfy the same receiving policy. The more informative observation about enforcement is within the SIS delegation cell: all six trials ultimately completed, yet 15 write attempts were rejected for scope, 10 for missing grants and 4 for the wrong entity. Across those trials, 7 grants covered 8 admitted effects. Completion and enforcement can therefore coexist in one workflow, and the attempted/admitted distinction has an observable consequence. These trials characterize destination gatekeeping under the tested transfer protocol.

Reviewer_ziuUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Skill-in-Skill (SIS), a portable agent-skill format that binds a capability KK to execution constraints �1��1� and governance rules �2��2�. TT specifies task applicability, �4��4� specifies execution workflow constraints, �5��5� specifies required runtime evidence, ss specifies scope, �7��7� specifies authorization, and vv specifies verification. At a destination, SIS evaluates task applicability, scope, and authorization before activation, issues a local execution grant, records an evidence chain, and verifies the resulting terminal state.

The paper compares SIS with an External Governance Sidecar on 120 frozen business workflow instances across finance, support, legal, operations, and sales. Each task is run three times under each condition using three model routes: DeepSeek-V4.1-Flash, MiMo-V2.6-Pro, and Qwen3.8-27B. The reported results show reductions in file reads of approximately 7.6–11.2%, tool-call reductions of 2.6–9.1%, and a maximum reported wall-clock reduction of 13.2% for Qwen. Repeated task-context reads decrease by 76.4–90.3%. Verified completion remains near ceiling. In a small transfer experiment, SIS achieves 100% authorized completion after site rebinding and principal delegation, while SIDECAR achieves 0%; both are blocked when destination governance is withheld.

The main contribution is an architectural proposal and controlled proof-of-concept evaluation rather than a new learning algorithm or formally established security protocol.

Critical strengths
  • Clear problem decomposition and practical framing. The paper identifies a real systems concern: capability artifacts and their valid-use conditions can become separated during transfer. The decomposition into applicability, scope, authorization, workflow constraints, evidence, and verification is concrete enough to be useful as a design vocabulary.

  • More informative instrumentation than terminal accuracy alone. The paired comparison records file reads, tool calls, token usage, latency, and verified effects, and it includes an independent terminal-state check and within-suite held-out instances. This is a stronger evaluation structure than reporting only aggregate task success.

Critical weaknesses
  • The technical novelty and security mechanism are not established. SIS is presented as a capability bundled with constraints and a governance manifest, but the manuscript does not define a new enforcement algorithm, formal semantics, compositional security property, trust model, or proof. Enforcement still depends on the destination runtime. The paper therefore appears to repackage established policy-bearing and proof-carrying artifact ideas without demonstrating a decisive technical distinction.

  • The main evaluation is confounded and too narrow to support the claims. SIDECAR is denied destination-side reassociation in the transfer experiment, making the 0% versus 100% result largely protocol-induced. The efficiency comparison lacks a cached/precompiled/equally loaded sidecar control. All experiments use one harness and synthetic fixed workflows, with no independent sites, real cross-runtime transfer, adversarial behavior, open-ended skill composition, or external OOD benchmark.

Questions for authors

Clarification requests

  1. What is the exact fair SIDECAR baseline? Please provide the complete SIS and SIDECAR manifests, loading paths, prompts, cache policy, and destination protocol. What happens when the sidecar manifest is transferred with the capability and explicitly re-associated or precompiled at the destination? Does SIS still reduce reads, tool calls, tokens, and latency under that control?

  2. What does SIS actually guarantee when the destination runtime is trusted but the package, agent, or host may be compromised? Please specify the signer and trust-root model, identity and delegation mechanism, key rotation, revocation, expiration, freshness, replay and downgrade prevention, dependency closure, policy conflicts, and nested-skill semantics. Which guarantees come from the package itself, and which come exclusively from the destination runtime?

  3. What concrete heterogeneous environments were tested? “Site rebinding” and “principal delegation” appear to be regimes inside one DeepSeek Harness/local-process setup. Were any independent runtimes, policy engines, communication interfaces, identity providers, or administrative domains used?

  4. Can the authors release the implementation and raw evidence needed for independent reproduction? This should include the package format, task fixtures, verifier, raw traces, model route versions, sampling parameters, seeds, hardware, API configuration, and exact cache behavior.

  5. How should Table 5 be interpreted statistically? The table states six terminal trials per event but reports 13 grants, 14 covered effects, and 29 rejected writes. Please provide the trial-level accounting, denominators, confidence intervals, and the relationship between terminal outcomes and effect-level events.

  6. How were repeated trials and multiple endpoints handled? The manuscript averages three repeats before bootstrapping over fixed task instances. Please justify this choice, report within-task stochastic variability, and explain whether any correction or preregistered strategy was used for the large number of task-class and metric comparisons.

Suggestions for future work

  • Evaluate tampered or malicious skills, prompt injection, compromised agents, revoked credentials, stale policies, replayed grants, dependency substitution, policy conflicts, and nested/composed skills.
  • Compare against well-engineered external governance systems with caching, local compilation, explicit destination reassociation, signed manifests, and equivalent initial context.
  • Test real heterogeneous runtimes and independent domains, together with open-ended agent tasks rather than only fixed read–transform–submit–verify workflows.
  • Report actual end-to-end cost, including package transfer, initial loading, context tokens, latency variance, and monetary or energy cost, rather than relying primarily on file-read and tool-call counts.
1 comment

The contribution is a skill format that keeps a capability associated with the conditions and verifier a receiver needs to admit its effects. Three decisions make this concrete: the package identifies the binding, the destination issues authority, and an independent verifier checks the outcome. Your questions help specify the evidence for each decision.

1. Baseline and loading. We checked the staged files and rendered prompts for a matched pair of formal MiMo trials. Both placements have separate governance.json and binding.json files. Their directories are stage/sidecar/ for SIDECAR and stage/_sis/ for SIS. Both prompts say “Read and follow the governance and binding at” and give the respective binding path. Across the suite, the observed differences are 0.278–0.475 fewer file reads and 0.269–0.933 fewer tool calls per run. 2. Trust and the contribution over policy-bearing artifacts. The destination maintains the expected capability, governance and verifier digests independently of the arriving declaration. It validates the actor and scope and, for a delegated principal, requires a delegation record before granting access. An agent’s request or a package’s assertion cannot widen the scope the receiver admits. The trusted components are the local policy, the enforcement path mediating writes and the terminal verifier. A host compromise that subverts those components is outside this model. The artifact carries an association that can be checked on arrival. The grant and mediation of effects come from the destination. Appendix G describes signed delegation; the submission establishes no general package certification hierarchy or key-lifecycle protocol covering rotation, revocation, expiry, freshness, replay or downgrade. Likewise, the bound digests cover the named artifacts; dependency closure, nested-skill composition and general policy conflict resolution require additional semantics. We accordingly frame the advance as a portable unit of admission for agent skills and an instrumented evaluation of that unit.

3. Heterogeneity. The three hosted routes share DSH and one local-process harness. The DeepSeek boundary experiment changes the site-specific binding or delegated principal. The 40 held-out tasks are within-suite instances.

4. Reproduction. Table 7 gives provider routes; Appendix B gives the 90-request cap, 32,768-output-token cap and one retry. The archived MiMo pair above contains the staged files, actual task prompts, runner specifications and execution records. Those artifacts make the loading-path comparison inspectable. We cannot confirm a public anonymous release link here.

5. Table 5 denominators. Panel (a) has six terminal trials per placement and event. Panel (b) counts operations during the twelve SIS trials with destination governance available.

SIS eventCompleted trialsGrants issuedCovered effectsRejected writes
Site rebinding6/6660
Principal delegation6/67829

The delegation record further resolves the 29 writes into 15 scope denials, 10 attempts without a grant and 4 attempts on the wrong entity. Rejected attempts can precede an authorized completion in the same trial. Those 29 observations share six trials and cannot serve as independent terminal observations. For scale, the two-sided 95% exact binomial intervals for 6/6 and 0/6 are [54.1%, 100%] and [0%, 45.9%], respectively, under an independent-trial interpretation. The summary supports the event breakdown above; a per-trial allocation requires the individual ledgers.

6. Repeats and endpoints. Each task receives equal weight: we average three repeats within each task and placement, compute the paired difference, then bootstrap task instances over 10,000 resamples. The unit matches the question “how much does resource use change for a task in this suite?” It avoids counting three executions of the same input as three distinct tasks. These intervals describe task-level contrasts. The paper does not separately report within-task dispersion or specify preregistration or a multiplicity correction. We interpret the model-by-class and secondary endpoint intervals as pointwise descriptive results. For end-to-end interpretation, the directly measured Qwen wall-clock saving is 8.29 seconds per run; converting read counts to monetary or energy savings would require further measurements.

Reviewer_eDhLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Skill-in-Skill (SIS), which packages a reusable capability together with execution constraints and governance rules. SIS uses destination-side authorization and independent verification to support governed execution across environments. Experiments show that SIS reduces file reads and tool calls compared with an External Governance Sidecar (SIDECAR), while largely preserving verified completion.

Critical strengths
  • The paper addresses an important problem that can arise when skills are transferred across execution environments.
  • Across 120 task instances and three models, SIS consistently reduces file reads and tool calls compared with SIDECAR.
  • The evaluation includes paired comparisons, confidence intervals, and held-out instances, while largely preserving verified completion.
Critical weaknesses
  • The comparison does not include a SIDECAR design that transfers an external governance specification alongside the capability and re-associates them at the destination, or one that uses caching or preloading to reduce repeated context access.
  • In the cross-boundary transfer experiments, SIDECAR lacks the governance binding required at the destination. Thus, the reported 100% authorized completion for SIS versus 0% for SIDECAR does not fully establish an advantage over stronger alternatives.
  • The evidence for governance portability across execution environments remains limited. All three models are evaluated using the same DeepSeek Harness (DSH), and the task categories share a similar read–transform–submit–verify workflow. Although the paper evaluates site rebinding and principal delegation, it does not sufficiently examine how differences or conflicts between governance rules at the source and destination would be handled, or the associated reconfiguration cost.
Questions for authors
  1. Would SIS retain its efficiency and governance portability advantages if SIDECAR transferred an external governance specification alongside the capability and used caching or preloading?
  2. How does SIS handle conflicting governance rules between source and destination environments? What reconfiguration is required?
  3. Can the authors provide cross-boundary transfer results using different agent runtimes or different destination-side authorization mechanisms?
1 comment

Your question about transferring the sidecar identifies two operations that have different consequences: loading a contract into the agent’s context and establishing its validity at the destination.

Would preloading or transferring the sidecar close the gap? Preloading could recover some or all of the lookup savings. In a matched pair of archived MiMo runs, both initial prompts explicitly directed the agent to a binding file. SIDECAR placed it at stage/sidecar/binding.json; SIS placed it at stage/_sis/binding.json. Across the 120 paired tasks, SIS saved an average of 0.28 reads per MiMo run. On Qwen the corresponding difference was 0.48 reads, and elapsed time fell by 8.29 seconds per run (95% interval 4.52–12.09 seconds saved). These estimates apply to the measured loading paths. An equally preloaded comparison would need to count initial loading as well as subsequent lookups.

For transfer, a sidecar that delivers the same binding and successfully reassociates it can in principle obtain the same grant. Table 5 answers a narrower but practical question: what arrives when the skill itself is the unit of transfer? SIS delivered the binding needed for all six site-rebinding trials and all six delegation trials to complete. The external binding stayed at the origin under the tested SIDECAR protocol. The experiment identifies the provisioning step an external design has to supply; it leaves the time and effort for that step open.

What happens when the rules change? In our rebinding fixture the skill moves from account A to account B, with a revised capability and a B-specific contract. B checks the capability, governance and verifier digests against its own expected values, checks the actor, and requires the requested scope to fit B’s policy. Delegation additionally requires a local delegation record for the new principal. Thus a source permission for A cannot authorize a write to B. Reconfiguration consists of establishing the B-specific binding and, when the actor changes, its delegated authority. An incompatible request is blocked before effect. We measured the resulting admission decisions; administrator time and automatic reconciliation of conflicting policies remain open.

Which environments were tested? All three model routes use DSH and the same local-process harness. The boundary study uses DeepSeek and varies the destination binding and principal within that harness. The evidence therefore supports rebinding and delegation under one destination implementation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.