Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents

TL;DR

Agent Data Protocol (ADP) is a lightweight “interlingua” schema for agent interaction trajectories — a representation language designed to bridge heterogeneous agent datasets (API/tool-use, browsing, coding, software engineering) and unified agent SFT pipelines. The authors unify 13 existing agent training datasets into ADP, convert the standardized data into training-ready formats for multiple agent frameworks, and run SFT against the unified pool. The headline empirical result is an average ~20% gain over the corresponding base models and SOTA-or-near-SOTA performance on standard coding / browsing / tool-use / research benchmarks with no domain-specific tuning. Accepted to ICLR 2026 (Oral).

Key claims

The bottleneck for public-scale agent SFT is fragmentation across heterogeneous formats, tools, and interfaces — not lack of underlying data [Abstract].
ADP is expressive enough to cover API/tool use, browsing, coding, software engineering, and general agentic workflows in a single schema, while remaining simple to parse and train on without per-dataset engineering [Abstract].
The authors unified 13 existing agent training datasets into ADP format and converted the standardized ADP data into training-ready formats for multiple agent frameworks [Abstract].
Supervised finetuning on the unified ADP-formatted pool delivers an average ~20% gain over corresponding base models [Abstract].
The trained models reach state-of-the-art or near-SOTA performance on standard coding, browsing, tool use, and research benchmarks without domain-specific tuning [Abstract].
All code and data are released publicly to lower the barrier to standardized, scalable, reproducible agent training [Abstract].

Method

ADP is a lightweight representation language that captures the structure shared across agent trajectories: actions (tool calls, code generation), observations (environment state, tool responses, feedback), and reasoning traces (planning, decision rationale). Existing datasets are converted via per-dataset adapters into ADP once; downstream pipelines then read ADP and emit framework-specific training pairs. The project frames this as a many-to-many → one-to-many engineering shift: instead of N datasets × M agent frameworks of bespoke conversion code, the team writes N adapters into ADP plus M adapters out, eliminating redundant per-pair engineering across projects.

The validation experiment unifies 13 agent training datasets spanning API/tool use, browsing, coding, and SWE, then runs SFT on the merged ADP-formatted pool. Reported results: ~20% average gain over base models across standard benchmarks (coding, browsing, tool use, research), with the same recipe reaching SOTA or near-SOTA on each without domain-specific tuning. The project page lists “Action Tracking”, “Observation Capture”, “Reasoning Traces”, and “Quality Assurance and Analysis” as the first-class schema features.

Results

Unification scope: 13 existing agent training datasets converted to ADP and made training-ready for multiple agent frameworks [Abstract].
Headline gain: ~20% average improvement over corresponding base models after SFT on the unified ADP pool [Abstract].
Benchmark posture: SOTA or near-SOTA on standard coding / browsing / tool-use / research benchmarks without domain-specific tuning [Abstract].
Venue: ICLR 2026 Oral.

Per-benchmark numbers and per-dataset ablations are in the OpenReview submission (tG6301ORHd) rather than the project landing page; the page itself surfaces only the headline framing.

Why it’s interesting

ADP is the data-side counterpart to the recipes filed under Tool-Use Agents and Agentic Software Engineering. Those concept pages already document the convergence on multi-stage post-training (SFT → DPO → GRPO) and the load-bearing role of verifiable executable tasks — see Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing for closed-loop tool-trajectory synthesis (32B base 19.8 → 70.9 on BFCL via SFT+RL on synthesized trajectories) and Qwen3-Coder-Next Technical Report for ~800K verifiable SWE tasks mined from GitHub PRs with executable Docker verifiers. ADP attacks the orthogonal question: given that you already have 13 such datasets in the wild, how do you stop rewriting the data-loader for each one? It proposes a schema, not a new training recipe, and validates by showing the unified pool alone gets ~20% over base across the standard benchmark quadruple.

Within Synthetic Training Data, ADP is closest in spirit to the “curation gate / pipeline-quality is load-bearing” thesis but acts at a different layer — instead of generating new data via multi-agent role-play (Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing) or hindsight curation (Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning), it standardizes the format of already-collected data so heterogeneous sources can be SFT’d together without per-source engineering. This complements MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling‘s parameter-space recipe (large synthesized trajectory corpora + multi-stage post-training on a single model) by reducing the per-dataset friction that currently forces each lab to re-invent its own trajectory loader. The Carnegie Mellon / OpenHands authorship overlap with prior agent-training infrastructure (OpenHands, SWE-Bench-style execution environments) makes ADP a strong candidate for being adopted as the de-facto interchange format if the released adapters cover the standard 13 datasets cleanly.