Multi-Tool Composition
Almost nothing real is one tool call. Finding last quarter's unpaid invoices and sending finance a summary needs a search, a read of each result, some arithmetic and a send — and every step's output has to arrive intact at the next. Multi-tool composition is the agent's ability to chain calls into a sequence that actually works.
How calls get composed
Three patterns cover most of it:
- Sequential threading: a value from one result — a record id, a file path, a cursor — is copied into the next call's arguments. The model must reproduce it exactly, which is where a small slip becomes the wrong record.
- Parallel fan-out and join: independent calls issued together and merged once all have returned, cutting the round trips a long chain would cost.
- Code as composition: given a code-execution tool, the model writes a short script that calls several tools, loops and filters — one step instead of twenty, and the intermediate data never enters the context window.
Why it matters
Composition is where tool use stops being a demo and starts being work, and its failures are specific. An id that scrolled out of the context window gets plausibly invented instead of looked up. Two schemas disagree about units or date format, and the chain completes with a wrong number. A break halfway leaves part of the work applied and nothing to roll it back. Strict schemas, compact tool results, and a habit of re-reading state rather than trusting memory are what keep long chains honest.
Multi-tool composition is an agent chaining several tool calls so that each result feeds the next, completing work no single function covers — search, read, calculate, send. The chain may also fan out in parallel and merge, or be expressed as a short script the model writes. It matters because composition is where a plan becomes executed work, and where one mis-copied identifier or mismatched schema yields a confidently wrong final answer.
Think of someone cooking from a recipe. They don't use one gadget — they reach for the peeler, then the pan, then the oven, and each step depends on the last one being done properly. Multi-tool composition is software doing the same thing. Given a goal, it picks one tool, looks at what came back, and uses that result to decide which tool to reach for next.
This is what lets a program handle a real errand rather than a single lookup. The catch is the same as in the kitchen: if an early step goes wrong, everything after it is built on a mistake.