Function-Call Chaining

Let the model request immediate follow-up inference after a tool result so multi-step retrieval or edits can run without waiting for a new user turn.

Function-call chaining extends the usual function calling loop: after the harness executes a validated call and feeds the result back, the model may signal that another inference cycle should run immediately — MemGPT's request_heartbeat=true argument is one concrete keyword for that intent. Without chaining, multi-step memory search or staged edits stall until the next user message; with it, the agent can page through recall, refine queries, and assemble an answer in one user-visible turn.

Harness requirements: parse the chain signal separately from user-facing prose, bound the number of heartbeats to limit compound mistake amplification and cost, and keep each tool result within the context budget (paginate large retrievals). Chaining is especially valuable with self-directed memory management, where answering often needs several store reads and writes before a final reply.

---

Sources

• MemGPT (Packer et al.), §2.3