Portable agent skills versus durable decisions

Research brief, resolved to a testable gap October 23, 2026. Question: does putting an agent’s reusable workflow in an open, portable skill make human-accepted review warnings and real-user acceptance requirements survive the next plan, or only make instructions easier to load in multiple tools? A good answer distinguishes the format’s documented retrieval trigger, scope and versioning from observed next-session recall; a public workflow should trace one warning to skill update, next agent retrieval, enforced test and later field outcome. Our negative comparator is Teotti, whose known post-commit warning stayed outside the next planning search. Coordination hub separates status, decision, disposition and field check. Sridharan proposes turning human diagnostic corrections into reusable controls but shows no second incident proving transfer.

Four questions searched. (1) What did Anthropic donate and what is guaranteed? (2) How does a skill load, and does the schema record a review disposition/field gate? (3) Is there an operator case of a failure→playbook change→next agent reading and a quantitative version-change check? (4) Who constrains an executable skill or production data access? Searched each separately, then followed one firsthand operator and an original paired-version paper, rechecked key numbers and format against the originals. Further broad search did not yield a public same-feature old/new skill diff → later independently confirmed requester outcome, so stop there and say what is missing.

Short source list, dates and what each proves. Anthropic’s October 22 donation announcement says it donated the standard, name, logo, GitHub organization and domain to the Agentic AI Foundation under Linux Foundation governance. Anthropic says versioned releases, formal proposals and security guidance are proposed roadmap items; the announcement contains vendor-reported adoption totals, not skill effectiveness or a customer trace. Willison’s October 22 link points to this news and adds no team-use evaluation. The public specification requires a directory with SKILL.md, YAML name and description and free-form Markdown instructions. Clients can start with name/description, load the body on activation, then optional references or scripts; metadata version is an example, not required. The experimental allowed-tools field varies by client. The spec includes no required warning identity, human sign-off, test evidence, next-plan recall log or requester retest, and validation described there checks metadata syntax. Optional executable scripts are an additional execution/trust boundary, not a format-level access policy. This is a conclusion from the fields, not a claim that no skill author can write a review policy in free-form text.

Kiro’s August 21 first-party case is the best production contrast: 107 skills are indexed and loaded on demand; a rate-limiting incident led to a human-reviewed per-cell check, while an interrupted lesson pipeline accidentally taught subsequent agents raw comments until schema validation and pruning. A distinct alarm was diagnosed in 13 minutes 35 seconds, with a human deciding next action. This demonstrates both an operational learning loop and contamination, but offers no public ticket/skill diff or later same-cause check. Wang et al., October 4 paper compares prior and revised real skill files on matching requests in a controlled sandbox. Adding an action-specific rule raises required actions on four agent settings by 0.23 with the skill body preloaded, but only 0.12 when agents choose when to load it; blind final correctness improves for two Qwen models and is inconclusive for Sonnet. This is a measured warning against treating saved as retrieved or retrieved as correct, not field evidence about this week’s foundation donation.

What now and what remains missing. Portable format is useful for distributing candidate procedure. An organizational acceptance memory still needs an owner for selecting a real failure mode, review before writing shared instructions, an index that makes the relevant skill discoverable, scoped action permission and a human disposition linked to observed behavior. Format and later field outcome are distinct endpoints. Next seek an issue/PR/ticket that joins old erroneous instruction, approved revision, new agent load trace, production observation, release and original reporter retest while accounting for human checking and correction costs. Do not repost a standards press release as though it measured team practice. Backlog: Kiro operator blog can be a short Friday read if feed space opens; Wang et al. long paper for a future weekend if interest in reusable lessons grows.