TanStack AI adds hard caps on agent tool calls

TanStack shipped bounded tool-call limits in TanStack/ai, its provider-agnostic TypeScript SDK for OpenAI, Anthropic, Gemini, and Ollama, per the project's GitHub release.

The library has 2,953 stars and adapters for React, Vue, Svelte, and Solid. That reaches a wide slice of teams wiring tool calls into production agents. Before this, stopping a runaway loop meant writing a counter yourself and hoping you didn't miss an edge case.

AgentLoopState exposes toolCallCount and lastTurnToolCallCount so a running agent can report its own usage. The method maxToolCalls(n) caps the cumulative number of tool calls across a whole session. Separately, chat({ maxToolCallsPerTurn }) caps how many calls can run in parallel within a single turn.

That state sits inside the agent loop, not bolted onto the API client as a separate wrapper. A developer can inspect toolCallCount mid-run and choose to bail, warn, or throttle the remaining budget instead of hitting a wall with no warning.

The release notes don't say what happens when a call hits the cap mid-task. Does the loop return a partial result, throw, or truncate the last turn silently? That's worth checking before shipping this in anything user-facing, since a silent truncation on a multi-step task is worse than no cap at all.

Whether Vercel's AI SDK or LangChain.js add an equivalent cap next is worth watching. A runaway tool loop burning through a budget is a support ticket every team running tool-calling in production has probably already filed.