Anthropic Hands Agent Skills to an Independent Foundation, Making Interoperability the New Agent Fight

The most revealing part of Anthropic’s announcement is not a new benchmark or a larger context window. It is the decision to move Agent Skills, a format for packaging repeatable instructions and supporting files, toward an independent foundation. Anthropic announced the handoff on October 22, 2026; that announcement date matters because the move is about governance of a format already used in real workflows, not a promise about a future model. The primary account is Anthropic’s own announcement: the company’s description of the foundation move.

The file that matters more than the model

An Agent Skill is best understood as a small operational bundle. It can describe when an agent should use a capability, what files or references it needs, which tools are allowed, and how a human should review the result. That makes it closer to a runbook than to a prompt. The distinction matters. Prompts are usually copied into a conversation; skills are meant to travel with a workflow. Anthropic’s public Agent Skills repository gives builders a concrete view of that packaging model.

For Anthropic’s decision to place Agent Skills under an independent foundation, that distinction changes the operating question. The evidence has to be tied to the named system, its date, its owners, and the boundary between what the announcement confirms and what an operator must still verify. That is why this story belongs to practitioners as much as to headline readers.

Why Anthropic is giving up control of the format

Skills are instructions, not magic capabilities

The handoff also separates two things vendors have often blended together: the intelligence of a model and the portability of a task. A model may be excellent at drafting a contract summary, but an organization does not want the operational knowledge for that task trapped inside one chat product. A skill can describe the review sequence, source hierarchy, redaction rule, and escalation point. The model supplies judgment; the skill supplies a repeatable operating envelope.

Portability changes the economics of agent platforms

Anthropic has an incentive to make the format portable. Claude Code, Claude’s tool-use products, and the company’s broader agent strategy benefit when developers can bring workflows with them. The move lowers the cost of trying Claude without asking a team to rewrite every internal procedure. It is a distribution strategy, but it is also a standards strategy. The same logic has appeared in Anthropic’s support for Model Context Protocol, where a shared interface can increase the number of useful integrations around a model.

The security review moves into the skill package

That incentive does not make the standard fake. It makes the governance question sharper. If a vendor controls both the runtime and the package format, compatibility can quietly become a retention mechanism. If an independent steward controls the specification, a different failure becomes possible: the document can be neutral while the important behavior remains proprietary. A foundation must therefore publish conformance tests, version rules, and a clear explanation of what an implementation is allowed to ignore.

What an independent steward can and cannot fix

The practical value of a skill is visible in the boring details. A finance skill might require the agent to read a ledger export, reconcile dates in a specified timezone, cite each adjustment, and stop before posting a journal entry. A support skill might allow a refund recommendation but prohibit the refund action unless an employee approves it. Those details are more important than a glossy demonstration because they define the boundary between assistance and delegated authority.

The developer workflow after the handoff

The security problem follows the package. A skill can contain a malicious instruction, an unsafe tool description, a hidden data-exfiltration path, or a dependency that changes after approval. Anthropic’s own building-effective-agents guidance emphasizes simple, composable systems, but a portable skill makes composition a supply-chain issue. Teams will need signatures, provenance, reviewable diffs, and a way to pin a skill to known dependencies.

The enterprise buyer gets a new question

This is where the foundation model of governance matters. The NIST AI Risk Management Framework treats accountability, documentation, and monitoring as system properties. A skills standard can make those properties easier to express, but it cannot guarantee that an agent runtime will enforce them. A package may say ‘require approval’; only the runtime can prevent an unapproved API call.

Why open standards often arrive after a product wins

Interoperability also changes the commercial shape of agent platforms. Today, a consultancy can charge for rebuilding the same workflow for each model vendor. A common skill format could shift value toward domain expertise, evaluation data, connectors, and controls. That is good for buyers who want choice, but uncomfortable for vendors whose moat is a growing library of proprietary workflow templates. The prize moves from owning the wrapper to owning the evidence that the wrapper works.

A standard is only useful when failure is portable

The neighboring standards show both the opportunity and the trap. MCP focuses on connecting models to tools and data. A2A focuses on agent-to-agent communication. OpenAI’s Responses API and other vendor interfaces define runtime behavior. Skills sit above those layers: they describe how a job should be performed. If the layers are confused, a skill becomes a bag of vendor-specific instructions. If they are separated, a team can change the model or connector while retaining the operating procedure.

The missing pieces are versioning and provenance

An independent foundation can help most by defining the boring contracts. It should specify metadata, capability declarations, input and output expectations, lifecycle events, approval hooks, and error semantics. It should also state what is not standardized. A short, testable specification is more valuable than an ambitious catalog of every possible agent behavior. The Agentic AI Foundation announcement is therefore only the beginning; stewardship has to become engineering work.

How to test a skill without trusting its author

Versioning is the first hard case. A skill can change without changing its name, especially when it points to external documentation or a live tool. An enterprise approval that covered one version cannot silently cover another. Teams should pin versions, record checksums, and review changes as they would review a dependency update. A package manager for skills may eventually be as important as the format itself.

What this means for MCP, A2A, and tool catalogs

Evaluation is the second hard case. A skill can be syntactically valid and operationally useless. A good test suite would include ordinary requests, ambiguous requests, adversarial documents, missing permissions, stale data, and a forced human escalation. The score should include not only task completion but also refusal quality, citation quality, action reversibility, and audit completeness. Anthropic’s Claude Code documentation points toward this kind of workflow because the useful unit is a verified change, not a fluent answer.

The risk of a foundation becoming a neutral logo

The handoff gives enterprise buyers a more precise procurement question. Instead of asking whether a vendor ‘supports agents,’ a buyer can ask whether the vendor imports a skill without rewriting it, exposes every tool invocation, respects its approval gates, and exports the resulting audit trail. A vendor that claims compatibility but cannot answer those questions is offering branding, not interoperability.

A practical adoption plan for teams

There is a human consequence too. Portable skills make organizational knowledge easier to preserve, but they also make it easier to codify bad habits. A workflow that quietly discriminates, over-collects data, or routes all exceptions to an exhausted team can become an artifact that spreads. The foundation should encourage human-readable packages and explicit ownership so that the skill is reviewable by the people affected by its decisions.

The uncomfortable lesson for model vendors

The strongest adoption path is narrow. A team should choose one reversible workflow, such as preparing a weekly operating report, and write down the inputs, tools, evidence requirements, and stop conditions. It should run the skill in shadow mode, compare the result with the existing process, and expose every deviation. Only after the failure modes are known should the team connect a write-capable tool.

Watch the package, not the press release

For model vendors, the handoff is a test of confidence. A vendor that believes its advantage comes from model quality and runtime reliability should welcome portable workflows. A vendor that depends on lock-in will make portability sound easy while keeping the important metadata private. The market will learn which is which through migration tests, not launch events.

The evidence chain readers should keep

The claims in this article are anchored near the relevant primary material rather than presented as anonymous consensus. Announcement dates are kept separate from publication dates where the source makes that distinction. Vendor descriptions are attributed as vendor claims. The unresolved questions are deliberate: a launch page can establish what a company says it built, but only deployment evidence can establish how it behaves under load, pressure, or failure.

flowchart LR
 A[Primary announcement] --> B[Independent test] --> C[Operational evidence]
 C --> D[Human decision]
 D --> E[Verified deployment or remediation]

Sources and primary records

The following records were consulted for names, dates, mechanisms, and context:

The useful question after reading about Anthropic’s decision to place Agent Skills under an independent foundation is not whether the announcement sounds ambitious. It is whether the next public artifact makes the claim easier to test.

The contract should travel with the work

A portable skill is most useful when it carries the assumptions that normally live in a senior employee’s head. It should say which source wins when two systems disagree, what a reviewer must inspect, how long evidence remains valid, and what to do when the requested action falls outside the tested range. Those are not decorative instructions. They are the conditions that make a workflow transferable without pretending that every organization has the same risk appetite. A skill that exports only a prompt exports the least important part of the process.

The foundation’s first public test

The first serious test for the new steward should be a migration test. Take a real skill, run it on two independent runtimes, and compare tool calls, refusal behavior, evidence, and final outputs. Document every incompatibility. If the format cannot survive that test, the foundation should narrow the specification rather than declare success. Open standards earn trust by making failure visible.

Buyers should negotiate the exit

Every enterprise contract for agent software should include an exit clause for skills and evidence. The buyer should be able to export the package, its version history, evaluation results, and audit records in a readable form. That requirement turns portability from a marketing claim into a procurement control. It also protects the organization when a vendor changes pricing, retires a model, or changes the permissions of a tool.

The handoff will matter if it changes what teams can safely move. The durable artifact is not Anthropic’s announcement; it is the skill that another runtime can execute, explain, and stop for the same reasons.