Agentic tool use is the mechanism that operationalises action in an agentic system: a language model serving as the reasoning core invokes external tools — APIs, databases, code interpreters and the like — so that it can access up-to-date knowledge through search, improve computational accuracy through programmatic execution, and perform real-world operations through external systems. Agentic Tool Use in Large Language Models: A Survey motivates it by what a standard model cannot do on its own: its knowledge is bounded by its training data, which can produce factual errors or hallucinations, and it struggles to access real-time information, execute precise computation, or interact reliably with external systems.
Usage
That survey organises the practice into three paradigms distinguished by where the optimisation signal comes from, and presents them side by side as complementary rather than as successive replacements.
- Prompting as plug-and-play. The model’s parameters stay frozen and tool-use behaviour is elicited through instructions, demonstrations and feedback from the environment. The survey subdivides it into interleaved reasoning and action (a loop that alternates deliberation with external operations, using each observation to guide the next decision), decoupled planning and execution (a structured plan is produced first and then carried out, reducing repeated deliberation at every step), and program-aided reasoning (the intermediate representation shifts from text to executable code, making the program executor part of the reasoning process rather than an auxiliary utility).
- Supervised tool learning. Tool-use patterns are encoded directly into model parameters using labelled or synthetic supervision, which the survey frames as a shift from prompt engineering to data engineering. Its three strands are self-supervised data generation, large-scale instruction tuning across broad API corpora, and process-oriented or alignment-focused tuning that optimises the reasoning leading to a tool call rather than only the call’s correctness.
- Reward-driven tool policy learning. Tool use is treated as a sequential decision problem optimised through environmental feedback rather than imitation. Its strands are strategic decision optimisation (when to invoke a tool versus rely on internal knowledge), end-to-end policy learning over multi-turn interaction, and holistic or multimodal frameworks that optimise an entire agent system — memory module and tool selector included — as one policy.
The same survey treats evaluation as a cross-cutting dimension with three levels that it says have co-evolved with the paradigms: tool usage correctness (can the model produce structurally valid calls, retrieve and select the right tool, and decide whether a tool is needed at all), task completion (can it finish an end-to-end task with tools as functional components of the solution), and tool-driven interaction (can it act reliably in dynamic, stateful environments where its actions change the external world, including under adversarial conditions).
When It Applies
- Applies where a task needs information, computation or effects that the model’s parameters cannot supply — current knowledge, exact arithmetic, or changes to an external system. The survey’s framing is that these are recurrent mismatches between what models are good at and what deployment requires, not incidental gaps.
- Assumes a tool surface the model can be made to address correctly. The survey’s evaluation section treats schema compliance, retrieval from a large candidate pool, and correct argument grounding as separate abilities that can each fail independently, which is why it separates diagnostic benchmarks from end-to-end ones.
- Fails in ways that differ by paradigm, which is the survey’s argument for keeping all three in view. Prompting is flexible and cheap to deploy but suffers higher latency and token cost and becomes brittle over long horizons. Supervised learning is more efficient and stable but depends on obtaining large-scale, high-quality trajectory data. Reward-driven learning addresses exploration and error recovery but faces credit assignment across long trajectories where an early mistake propagates.
- Fails, too, when it is applied where it is not needed. The survey notes that many earlier benchmarks implicitly assume a tool should be called, and reports work showing that unnecessary invocation can degrade overall performance even when the tool is available — which is why deciding not to use a tool is treated as part of correctness.
- Well established as a research area with a substantial benchmark ecosystem, but the survey is explicit that its own unifying claim — that production-grade systems will combine all three paradigms — is its authors’ synthesis rather than a measured result.