Changing the Tool Interface Can Change the Agent's Decisions

Moving an agent from structured tool calls to command-line operations can change its decisions even when the available actions stay the same. A successful translation of the tool definitions does not establish equivalent behavior.

Imagine a hypothetical retail agent searching for a replacement item that meets a customer's size, color and price requirements. One deployment exposes named tools with argument fields. Another exposes commands with flags. Both reach the same catalog and purchasing service under the same permissions.

The agent could search differently, select a different item or spend more attempts repairing commands. The relevant deployment question is whether the customer still receives an eligible replacement, with the same constraints respected.

Match the operations before comparing behavior

In MICA, Zhenning Shi, Liang Pang and Yongqi Li generate MCP-style and CLI-style representations from a shared action contract. Decoded requests pass through common validation and execution. Their study covers six models and three environments: household tasks, web shopping and science tasks.

The reported performance differences reverse across model and environment combinations. Changes also appear in action selection and recovery, beyond malformed calls. This is a controlled comparison of model-facing representations, not a general contest between production MCP servers and unrestricted shells.

The authors also introduce DART. After an eligible failure that leaves the environment unchanged, it selects and validates an explicit action, then repairs its interface expression while preserving the operation and arguments. Its evaluation uses Gemini-2.5-Flash; that scope matters when considering transfer to another model.

Preserve meaning through recovery

One published shopping example makes the repair boundary concrete. Command quoting removes literal apostrophes from a product option. DART corrects the expression while retaining the exact selected option. Merely producing a parseable command would be insufficient if the repair changed which item the customer receives.

For WTK, this sharpens our runtime-configuration question. A catalog package compiled for another target needs assessment of how the model interprets its interface, including the feedback that shapes its next choice.

Our browser-action review follows a request through execution and observation. MICA adds a useful comparison: keep the executable operations fixed and examine what changes when their presentation changes.

We would ask whether the target preserves requested identifiers and constraints, whether retries drift into different actions, and whether successful outcomes justify the total interaction cost. A permitted, well-formed action can still be the wrong choice for the goal.

People can define those outcome requirements and review deployment changes while routine calls proceed autonomously. The assurance question belongs at the level of the compiled agent's behavior, not just the syntax its adapter accepts.

Source: Shi, Pang and Li, Matched Capabilities, Different Behaviors, arXiv:2610.34298v1, October 25, 2026, Sections 4-6 and Appendices B-E. The retail deployment example is ours.