Meta’s Fundamental AI Research division has derived a formal scaling law for tool-use agent training data — establishing, across 849 trained models and 458,752 generated tasks, that expanding the variety of APIs and database schemas an agent trains on produces larger generalization gains than generating more tasks on a fixed toolset. The finding inverts the default assumption behind most agent training pipelines, where data budgets are spent on task volume rather than tool breadth. For engineering teams building agentic systems, the implication is concrete: past a modest data budget, adding a new API to your training environment outperforms adding thousands of new tasks against APIs the agent has already seen.

The paper, submitted to arXiv on October 27, 2026, introduces SynthAPI — a controllable synthetic environment pipeline — and uses it to formally characterize the relationship between data composition and agent generalization. In doing so, Meta FAIR is attempting something for agent training data that Hoffmann et al.’s Chinchilla laws did for pretraining compute: replace ablation-based guesswork about data mix with a predictive equation labs can use to allocate a fixed budget before running a single experiment.

SynthAPI Generates Synthetic Environments, Not Just Tasks

Most agent training datasets are collected or synthetically generated as libraries of tasks: sequences of API calls, tool invocations, and database queries that demonstrate how an agent should handle a given goal. What distinguishes SynthAPI is that it treats the environment itself — the tools available, the databases populated, and the tasks drawn from them — as independently tunable training variables rather than a fixed backdrop.

The pipeline controls diversity along three axes simultaneously. The first axis is the set of APIs the agent can call during a task: SynthAPI can hold this fixed or expand it systematically. The second is the tasks drawn from those APIs: again, independently scalable. The third is the underlying database schemas the APIs operate on, determining what data the agent encounters even when the API structure is constant. By running training experiments that vary which dimension drives generalization while holding the others fixed, the authors can isolate exactly which aspect of diversity produces gains and at what scale.

Deriving a Scaling Law Across Four Model Sizes and 849 Training Runs

To produce a scaling law rather than an ablation, the Meta FAIR team ran what amounts to a systematic survey of the space: 849 models trained across four parameter scales, with each run varying the composition of the 458,752-task training dataset along the three diversity axes. Performance on held-out benchmarks was then mapped as a function of training composition, allowing the team to fit a predictive equation.

The resulting law formalizes what the ablations already suggest but could not previously quantify: the marginal return from adding more tasks on a seen toolset diminishes sharply as training budget grows, while the marginal return from expanding the tool and database variety holds up. At small budgets, task variety matters. At larger ones, API breadth is the dominant factor. The paper reports the law holds for models up to 180 billion parameters, giving it practical coverage across the model sizes most labs are deploying in production.

Why Agents Trained on Narrow Toolsets Fail in Deployment

The mechanistic explanation behind the scaling finding is not mysterious. An agent trained on a narrow set of APIs learns surface-level patterns: the specific argument structures, return types, and naming conventions of the tools it has seen. When deployed against a novel API in production — one with a different schema, different argument names, or different error behavior — the agent lacks the general principles of how to reason about tool interfaces because it has only memorized specific ones.

Expanding API diversity during training forces the model to develop higher-level representations of what an API is and how to interact with one, rather than memorizing what a specific set of tools looks like. Database schema diversity compounds this: even when the API surface stays constant, varying the underlying data the agent must query prevents the model from overfitting to training data structures encountered during training.

This failure mode is directly observable in the SynthAPI benchmarks. Agents trained on NVIDIA’s Nemotron-Agentic dataset — a large, widely used tool-use training dataset released in 2025 — underperformed expectations on multiple evaluation settings, including, in multiple cases, falling below the zero-shot baseline for base models that received no agent fine-tuning at all. The paper’s direct characterization attributes this to Nemotron-Agentic’s training data being “relatively simple and short-horizon data that does not generalize to the complexity of these benchmarks.”

Benchmark Results: What SynthAPI-Big Actually Achieves

Agents trained on SynthAPI data outperformed those trained on Nemotron-Agentic by a 30.5 percentage point advantage on average across three external evaluation benchmarks: AppWorld, the Berkeley Function Calling Leaderboard (BFCL/BFCLv3), and ASTRA-Bench.

On AppWorld Test-Challenge accuracy scores — a benchmark that evaluates agents on realistic multi-step task completion across a suite of simulated apps — SynthAPI-Big achieves 66.4% accuracy at the 4-billion-parameter scale and 77.0% at the 9-billion-parameter scale. Both scores exceed the prior SimSAT baseline by 7.6 and 11.6 percentage points respectively. On BFCL and ASTRA-Bench, SynthAPI-trained agents outperform the strongest per-benchmark baselines by 2.0 to 3.7 percentage points at their respective model sizes.

The paper uses four evaluation settings in total: AppWorld, BFCL, ASTRA-Bench, and a held-out SynthAPI-Holdout set. On SynthAPI-Holdout benchmark results, SynthAPI-Big achieves scores of 50.5 and 55.6. The SynthAPI-Holdout results validate that the gains are not an artifact of training and evaluation data drawn from similar distributions.

Independent Research Groups Reached the Same Conclusion in 2026

Meta FAIR is not the first team to report that diversity outperforms quantity for agent training data; the SynthAPI paper arrives as the third major 2026 result supporting this direction.

In February 2026, a team from the Shenzhen Institute of Advanced Technology, Alibaba, and UNSW Sydney published the TDScaling diversity scaling study (arXiv:2602.03219). That paper found that “under a fixed training budget, increasing trajectory diversity yields larger gains than adding more trajectories” — a finding derived independently, using a different agent type and different benchmarks.

In March 2026, researchers from Fudan University and MiniMax published the DIVE ICML 2026 paper (arXiv:2603.11076), accepted at ICML 2026, which reported that “diversity scaling consistently outperforms quantity scaling, even with 4× less data.” DIVE examined agentic task synthesis specifically for generalizable tool use — the closest prior work to SynthAPI’s domain — and reached identical directional conclusions.

What SynthAPI adds to this picture is the formal scaling law: a predictive equation rather than an ablation finding, derived from a sample size (849 trained models) large enough to support the mathematical formalization. The three papers together establish diversity-over-quantity as one of the more reliably replicated findings in the 2026 agent-training literature.

What the Scaling Law Actually Provides Labs

The practical value of a scaling law over an ablation is predictive power before the experiment. A lab deciding how to generate training data for a new agentic system can now use the SynthAPI law to estimate, for a given compute budget, what mix of API diversity and task count will maximize generalization — without running hundreds of exploratory training runs first.

This is structurally analogous to what Chinchilla optimal compute laws provided for pretraining: a way to treat data-parameter tradeoffs as an optimization problem with a known solution rather than an empirical guessing game. The SynthAPI paper makes this analogy explicit: the goal is a “compute-optimal” data composition strategy for agent training, parallel to compute-optimal parameter-token allocation in pretraining.

For organizations currently investing in agent training pipelines, the immediate implication is that infrastructure for generating diverse synthetic environments — not just large task libraries — is the critical bottleneck. SynthAPI is described in the paper as a reusable pipeline intended to support this, though as of October 29, 2026, no confirmed public release date has been announced.

Frequently Asked Questions

What is a scaling law for AI agent training, and why does this one matter?

A scaling law is a mathematical equation that predicts model performance as a function of some variable — in pretraining, usually the number of parameters or tokens. For agent training, the SynthAPI scaling law predicts generalization performance as a function of training data composition: specifically, how much of the data budget is spent on API diversity versus task volume. What makes this useful is that it is predictive rather than descriptive — a lab can use it to calculate the optimal data mix for a given compute budget before running any training, rather than discovering it through expensive ablation experiments, as described in the SynthAPI scaling law paper.

Why do agents trained on large datasets sometimes perform worse than untrained base models?

An agent trained on a large but narrow dataset can learn to pattern-match against the specific tool schemas and task structures in that dataset, effectively memorizing surface features rather than developing general reasoning about how to use tools. When evaluated on novel tools or databases not present in training, those memorized patterns transfer poorly, and the model can underperform a base model that had no such misleading priors to overcome. This is precisely what SynthAPI Nemotron comparison benchmarks observed on AppWorld and ASTRA-Bench evaluations.

How does SynthAPI control what kind of diversity goes into agent training data?

SynthAPI treats the training environment as three independently tunable variables: the set of APIs available to the agent, the tasks drawn from those APIs, and the database schemas the APIs operate on. Each can be scaled up or held constant during data generation. This means a team can run the same SynthAPI pipeline to produce a dataset optimized for API breadth, another optimized for task density, or any combination — and then use the scaling law to choose the mix appropriate to their compute budget, as specified in the SynthAPI design and pipeline paper.

Have other research groups confirmed that diversity beats quantity for agent training?

Yes. Two independent research teams published similar findings earlier in 2026. The TDScaling diversity beats quantity study (February 2026, arXiv:2602.03219), from researchers at SIAT, Alibaba, and UNSW Sydney, found that diversity of training trajectories outperforms raw trajectory count for code agents. The DIVE generalizable tool-use paper (March 2026, arXiv:2603.11076), from Fudan University and MiniMax, accepted at ICML 2026, found that diversity scaling outperforms quantity scaling even with four times less data for generalizable tool use. The SynthAPI paper extends both by deriving a formal predictive scaling law rather than an ablation finding.