agent-tool-integration

Safely expose read-only tools to LLM agents by injecting tool metadata into the system prompt and processing tool invocation via XML tags. This restricts agent capabilities to read-only data access without modifying core LLM invocation logic.

Restrict tool access in multi-agent workflows by applying a role-based tool allowlist. This pattern ensures specific agent roles only discover and invoke tools explicitly assigned to them from the local tool catalog.

Inject tool metadata into a system prompt by describing available tools, their side effects, and example inputs directly within the prompt context. This enables agents to discover local catalog capabilities without hardcoding tool names.

Agents can signal tool invocation using portable XML tags across different LLM providers. A post-processing tool execution handler parses these tags to deterministically execute the requested read-only tools.

Agent-tool-integration does not require modifying core LLM invocation logic. It bridges the gap between agents and external tools entirely through system-prompt injection and post-processing tool execution handlers.

Avoid system-prompt injection for tool integration when agents require write access or side-effect execution. This pattern is strictly scoped to read-only capabilities for fact-checking, data enrichment, and safe publishing workflows.