Action and integration

Tools connect model decisions to external observations and effects. Google's function-calling documentation explicitly assigns function execution to the application, while the model selects calls and arguments. This separation makes the tool boundary a place for validation and authority checks. gemini-function-calling

Contracts for model-selected actions

A useful contract identifies the operation, argument meaning, result meaning, errors and effects. Distinct names and bounded, informative results help the model choose and interpret tools; Anthropic describes evaluating these interface choices. tool-design

Schema-constrained output, illustrated by Gemini's structured-output support, constrains representation. It cannot establish that a selected account is the intended account or that a requested change is authorized. Check business invariants in execution code. gemini-structured-output

Recommended design separates transport failure, rejected input, failed operation and successful operation with an unexpected business outcome. Return actionable errors without leaking secrets. Preserve an operation identifier when a response is delayed or uncertain, so the caller can reconcile it.

Choosing the action surface

Surface	Useful property	Engineering obligation
Typed API	Explicit operation and data contract	Maintain schemas and authorization rules
Code or shell	Flexible composition and computation	Constrain filesystem, processes and network effects
Browser or GUI	Reach workflows without a suitable API	Observe state changes and recover from interface drift

SWE-agent demonstrates that a purpose-built agent-computer interface affects coding performance in its studied environment. It does not establish one best surface for every task. swe-agent

Evaluate realistic sequences, including malformed calls, large results and ambiguous outcomes. Screenshots and text outputs are observations, not proof that the desired state persisted. Tool-result truncation should be visible, with a route to the remaining evidence.

External integration

Interoperability addresses discovering and calling capabilities across implementations. Keep tool semantics independently testable even when a protocol transports them. Recovery of side effects belongs in durable execution; permission enforcement belongs in isolation and authorization.

References

• Function calling with the Gemini API — gemini-function-calling.
• Structured outputs — gemini-structured-output.
• Writing effective tools for AI agents—using AI agents — tool-design.
• SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering — swe-agent.