**DRAFT — pending editorial expansion.** This article is a working draft published as scaffolding for the NINtec content programme. The current version covers the substantive perspective in compressed form; the published version will expand each section to the 2,000+ word depth the topic warrants. Editorial review is required before promotion.
Tool use in Claude is the integration surface between Claude reasoning and external systems. Done well, tool use is the foundation of agentic systems. Done poorly, tool use produces deployments that break in subtle ways at scale.
Tool schema design
Tool definitions with input_schema (JSON Schema) tell Claude what arguments are valid. Underspecified schemas produce malformed tool calls. Overspecified schemas constrain Claude unnecessarily. The right balance is engineering work; we tune schemas against eval data.
Error semantics
Tools fail. Tool calls return errors that Claude must reason about. The error format, the retry semantics, and the escalation path all matter. Tools that fail silently produce agents that hallucinate workarounds. Production tools fail loudly with structured error information.
Parallel-call orchestration
Claude can issue parallel tool calls. The orchestration layer must handle concurrent execution, ordering dependencies between tools, and partial failure. Anti-pattern: assuming sequential semantics. Pattern: explicit parallelism with documented ordering.
Permission scoping
Destructive tool actions are gated. Claude can only invoke tools it has been granted; permission scoping is per-tenant and per-action. Anti-pattern: granting broad tool access early in deployment. Pattern: graduated grant of action authority based on observed behaviour.
Tool use done well is invisible in production — agents take the right actions, recover from failures, escalate when uncertain. Tool use done poorly produces incidents that take weeks to diagnose. The discipline matters.