Claude’s tool-use push makes interface design a reliability issue

Microsoft’s August 17, 2026 Foundry release added Claude tool search, helping models find interfaces without loading every definition upfront. Web fetch separately exposed domain restrictions and response limits. These details matter because an agent’s behavior depends on what an interface makes visible, not simply on the model’s quality. Five new Claude capabilities in Microsoft Foundry.

Calling a tool looks simple in a demonstration: the agent chooses a function and supplies arguments. In a real workflow, it also needs to know whether the function is appropriate, whether the arguments refer to the right object, and whether the response establishes success. Those decisions are the substance of tool-using autonomy.

Toolformer, submitted February 9, 2023, studied training a language model to select APIs, supply arguments, and incorporate results. Its examples included calculators and search tools. The research established a method for learning tool use, rather than a guarantee that arbitrary tools can be used safely without application design. Toolformer paper.

Test the boundary cases

Our recommendation is to build tool evaluations around ambiguity. Supply two similarly named records, an empty result, an expired identifier, and a response that is successful technically but incomplete operationally. These cases reveal whether the agent reasons about the tool's contract or simply reacts to familiar-looking text. A happy-path test cannot answer that question.

ReAct's October 6, 2022, paper linked reasoning with actions and subsequent observations. It offers a complementary perspective: a tool response should update the plan. Continuing as though nothing changed defeats the purpose of consulting the environment in the first place. ReAct paper.

Make outcomes legible

We would ask tool designers to return structured outcomes that distinguish completed work, partial work, and failure. The agent should preserve that distinction in its report. This makes retries more deliberate and escalation more useful. A robust tool interface does not eliminate uncertainty, but it gives the autonomous system clearer evidence on which to base its next decision.

Anthropic’s engineering account identifies the weak link

In a September 11, 2025 engineering article, Anthropic described improving tools by testing them with agents and inspecting failures. Its examples favored task-shaped interfaces over mechanically exposing every API endpoint, names that clarify purpose, and responses that preserve useful context without overwhelming the model. This was engineering guidance drawn from the company’s own work. It provides a more specific explanation of agent reliability than a general instruction to make the model reason harder. Anthropic writing effective tools.

A correct call can still be the wrong action

The Toolformer and ReAct research described below established useful ways for language models to choose external operations and incorporate their results. Enterprise deployment adds a less glamorous issue: the interface may be valid while the action is inappropriate. A customer search that returns the first matching surname is structurally successful, yet inadequate evidence for changing a subscription. The agent may not know that an omitted field changes the operation from a preview to a final submission unless the interface makes that distinction explicit.

This is why examples can add information that a schema lacks. A schema defines what a valid request looks like. An example can show which valid request belongs to which situation. That distinction is familiar in ordinary software documentation, but it takes on greater significance when a probabilistic model selects the request without a developer writing the exact call in advance.

The 2025 release supplies a concrete mechanism

Anthropic’s November 24 advanced-tool announcement introduced Tool Use Examples alongside search and programmatic calling. The company explained that examples can communicate optional-parameter combinations and usage conventions that JSON schemas do not express. In the same release, it noted that on-demand discovery introduces another step and is less advantageous for small catalogs of compact, frequently used tools. The product claim is therefore conditional: additional indirection can help large interfaces, while a simple application may gain little from it. Anthropic advanced tool use.

Tool quality is becoming a product differentiator

The Foundry announcement brings this engineering problem into platform competition. If an agent is sold as capable of acting across office and operational systems, then the clarity of its tools affects whether it can fulfill that promise. A confusing interface can waste a strong model’s capacity. A well-designed interface can make a modest task more predictable without changing the model’s weights.

There is a further implication for how failures are explained. “The AI made a mistake” is often too coarse a diagnosis. Did it choose the wrong operation? Did a search hide the relevant result? Did a response discard the record identifier? Did a default create an unintended side effect? Those questions point to different fixes. Treating every failure as a reasoning deficit can lead teams to buy a more expensive model while preserving the interface that caused the mistake.

The next phase of agent products may consequently reward companies that make their applications legible to automated operators. That does not require exposing unlimited functionality. It requires exposing the right distinctions. Tool search helps an agent find an interface; careful tool design helps it understand what using that interface will mean. Microsoft and Anthropic’s releases show both layers becoming explicit parts of the commercial stack, with their own limits and their own contribution to dependable autonomy.

The research behind the release

Toolformer, published in February 2023, trained a language model to insert useful API calls using a largely self-supervised procedure. Its tools included a calculator, search, translation, and a calendar. The research showed a path for learning when an external operation helps text generation. A commercial agent must add another layer: distinguish a useful operation from a permitted one, and a syntactically valid argument from the correct real-world target. That gap makes interface wording and examples operationally significant. They shape what the model believes the tool does before any returned result can correct the misunderstanding. Original research paper.