Agent Tool Design

Agent tool design covers the creation, selection, and execution of capabilities that an AI agent uses to interact with its environment. This includes defining functions, structuring skills, creating safe execution environments (sandboxing), and developing frameworks that govern how agents select and use tools.

Core Concepts

• Function Calling: A mechanism for LLMs to invoke external functions or tools.
• Model Context Protocol (MCP): An industry standard for agent-to-tool integration that enables dynamic service discovery through schema descriptions Formal Semantics for Agentic Tool Protocols: A Process Calculus Approach.
  • Formal analysis shows that while MCP is structurally similar to research frameworks like Schema-Guided Dialogue (SGD), it has expressivity gaps related to semantic completeness and explicit action boundaries Formal Semantics for Agentic Tool Protocols: A Process Calculus Approach.
  • A major practical issue in MCP is Description-Code Inconsistency (DCI), where a tool's natural language description does not faithfully reflect its underlying code implementation Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications.
  • A large-scale study of 2,214 real-world MCP servers found that 9.93% of description-code pairs exhibited inconsistencies, creating reliability and security risks Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications.

Tool Definition & Ergonomics

• Structured Skill Definition:
  • The Agent Instruction Protocol (AIP) models a skill as a directed execution graph in a schema-validated YAML, with nodes as discrete steps and explicit input/output edges AIP: A Graph Representation for Learning and Governing Agent Skills.
  • This structured approach avoids the unreliability of agents interpreting free-form prose and the fragility of editing prose-based skills AIP: A Graph Representation for Learning and Governing Agent Skills.
  • Compiling human-written skills to AIP increased the task pass rate for Claude Sonnet from 53% to 67% on SkillsBench, as it provides vetted, runnable units to the agent AIP: A Graph Representation for Learning and Governing Agent Skills.
• Self-Reflective APIs for Recovery:
  • Self-reflective APIs improve agent robustness by returning a machine-readable suggestions payload on validation failure, enabling the agent to repair its request and retry without external reasoning Self-Reflective APIs: Structure Beats Verbosity for AI Agent Recovery.
  • For Anthropic models, structured suggestions lifted task-completion rates by +36.7–40.0 percentage points over plain-English error messages Self-Reflective APIs: Structure Beats Verbosity for AI Agent Recovery.
• Specialized Tool Primitives:
  • MetaPoint enables precise spatial control in visual generation by representing a 2D coordinate as a single special token that the model interprets via its inherent positional encoding MetaPoint: Unlocking Precise Spatial Control in Agentic Visual Generation. This allows for pixel-level control without architectural changes.

Tool Selection & Orchestration

• Dynamic Skill Retrieval:
  • Static, task-level skill retrieval, where a fixed set of skills is chosen at the start, is often insufficient for dynamic web environments Online Skill Learning for Web Agents via State-Grounded Dynamic Retrieval.
  • State-Grounded Dynamic Retrieval (SGDR) is an online method that retrieves skills at each step based on both the overall task goal and the current webpage state, improving success rates on WebArena by over 10% Online Skill Learning for Web Agents via State-Grounded Dynamic Retrieval.
• Adaptive Tool Use:
  • The AgentMob framework uses an adaptive strategy for mobility prediction, employing a "fast path" for routine cases and triggering iterative, evidence-gathering tool use only for ambiguous situations Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent.
• Graph-Based Planning:
  • In domains with many tools, like bioinformatics, planning over a flat list of tool descriptions is inefficient Beyond Prompt-Based Planning: MCP-Native Graph Planning-based Biomedical Agent System.
  • The BioManus agent organizes a standardized MCP tool ecosystem into a typed graph of capabilities and plans over compact, task-specific subgraphs, decoupling planning complexity from the total number of tools Beyond Prompt-Based Planning: MCP-Native Graph Planning-based Biomedical Agent System.
• Reinforcement Learning for Tool Use:
  • Sci-PRM is a process reward model trained on scientific Chain-of-Tool trajectories to provide fine-grained supervision on tool selection, execution, and result interpretation SCI-PRM: A Tool Aware Process Reward Model for Scientific Reasoning Verification.

Execution and Safety

• Agent Harness:
  • The Microsoft Agent Framework provides an "agent harness" that integrates production-ready patterns for agent execution into a chat client Microsoft Agent Framework at BUILD 2026: Agent Harness, Hosted Agents, CodeAct, and more.
  • Key features include automatic context compaction to prevent overflow, session-scoped file memory, sandboxed shell execution, and human-in-the-loop approval flows for sensitive tools Microsoft Agent Framework at BUILD 2026: Agent Harness, Hosted Agents, CodeAct, and more.
• Sandboxing:
  • A controlled environment for code execution with restricted file system and network access.
• Skill Security & Governance:
  • Description-Code Inconsistency (DCI): This inconsistency between a tool's description and its implementation creates a critical defense blind spot, enabling risks from operational failures to malicious behavior Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications. The DCIChecker framework automates detection using static analysis and LLM-based validation Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications.

Reliability and Self-Healing

• Self-Correction via API Feedback:
  • Self-reflective APIs that provide structured, machine-readable suggestions on validation failure enable agents to reliably self-correct and retry failed tool calls Self-Reflective APIs: Structure Beats Verbosity for AI Agent Recovery.
• Constrained Generation:
  • Constrained Adaptive Rejection Sampling (CARS) is a technique to ensure that generated outputs, such as tool calls or code, satisfy strict syntactic or semantic constraints without distorting the LLM's probability distribution Constrained Adaptive Rejection Sampling.
  • It improves sample efficiency by adaptively recording and pruning constraint-violating token sequences in a trie so they are never revisited Constrained Adaptive Rejection Sampling.

Key References

• https://tldr.takara.ai/p/2606.05037
• https://tldr.takara.ai/p/2606.05031
• https://tldr.takara.ai/p/2606.04781
• https://arxiv.org/abs/2606.04769
• https://arxiv.org/abs/2606.04391
• https://arxiv.org/abs/2606.04494
• https://arxiv.org/abs/2606.05130
• https://arxiv.org/abs/2603.24747
• https://arxiv.org/abs/2606.04579
• https://arxiv.org/abs/2510.01902
• https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-at-build-2026-announce/
• https://arxiv.org/abs/2606.05037