Agent Tools
Agent tools are external functions, APIs, retrieval systems, user-interface controls, or specialist agents through which an AI agent obtains information or changes its environment.
Roles
OpenAI groups tools into three functional types. OpenAI source record
- Data tools retrieve context, such as records, documents, or search results.
- Action tools change external state, such as sending a message or updating a record.
- Orchestration tools expose specialist agents as callable capabilities.
Anthropic places retrieval, tools, and memory around an augmented language model and emphasizes that the model must receive environmental results as ground truth during execution. Anthropic source record
In ReAct, actions alternate with observations so that new evidence can update the next reasoning step. ReAct source record
Tools, protocols, and skills
A tool exposes a capability; an interoperability protocol standardizes how that capability is discovered or invoked; an agent skill packages procedural guidance for deciding how to perform recurring work. These layers can be combined but should not be conflated. Tools and interoperability source record (pp. 6–17) · Agent Skills source record (pp. 11–14)
MCP can reduce repeated model-to-tool connector code, while A2A addresses stateful coordination among agents. Anthropic’s launch source supports the original client–server motivation; neither protocol removes the need for clear semantics, scoped credentials, validation, and observability. Anthropic MCP announcement · Tools and interoperability source record (pp. 10–31)
Interface quality
Both practical guides argue that tool quality affects behavior, not merely developer convenience. Standardized definitions, descriptive names, clear parameters, examples, edge-case behavior, and explicit boundaries improve selection and reduce avoidable errors. OpenAI source record · Anthropic source record
Anthropic calls this design problem the agent–computer interface (ACI) and compares its importance to human–computer interface design. It recommends choosing formats that are natural for the model and avoiding unnecessary formatting burdens. Anthropic source record
OpenAI notes that overlapping tools can cause selection failures even when there are relatively few tools. It recommends improving tool clarity before splitting a system into more agents. OpenAI source record
Risk-aware design
Tool permissions should match the minimum capability needed for the task. OpenAI recommends rating tools by properties such as read versus write access, reversibility, required permissions, and financial impact, then using those ratings to trigger additional checks or human approval. OpenAI source record
See Agent safety and control for layered guardrails and escalation.
Analysis: tool contract checklist
The sources jointly imply that a production tool contract should specify:
- purpose and non-goals;
- parameter meanings and valid ranges;
- required permissions;
- success and failure outputs;
- side effects and reversibility;
- examples and edge cases;
- risk class and approval requirements;
- how observations are returned to the agent.
This checklist is synthesis, not a verbatim framework from either source.
Evidence status
The interface recommendations are consistent across several vendor guides and conceptually supported by ReAct’s use of observations. Confidence is medium because reliability still depends on model, task, protocol implementation, and deployment-specific testing.