Agentic Environment-Task Discovery

• Agentic environment-task discovery is a research domain that designs, synthesizes, and inverts environments to autonomously generate verifiable tasks for agent learning.
• It employs methods like API inventories, Docker inversion, and database formulations to create diverse, executable tasks with measurable success metrics.
• The field advances embodied science by coupling perception, language, action, and discovery, fostering robust, scalable autonomy in agent design and control.

Agentic environment-task discovery denotes a line of research in which the environment itself becomes an object of design, synthesis, inversion, or abstraction, so that agents can acquire capabilities through interaction rather than through fixed prompt-only supervision. Across recent work, environments are materialized from API inventories, grounded in real-world protocols, derived from healthy runtime systems by inversion, abstracted from controllable corpora, or explored as task-free sandboxes; tasks are then induced as executable trajectories, goal specifications, retrieval problems, or controller-synthesis episodes with explicit verification signals. A closely related scientific framing, “embodied science,” argues that discovery should be treated as a closed loop coupling agentic reasoning with physical execution through a Perception-Language-Action-Discovery framework, extending the same logic beyond simulators into empirical experimentation (Zhuang et al., 20 Mar 2026, Fang et al., 16 Sep 2025, Shi et al., 1 Jun 2026, Lin et al., 11 Feb 2026, Zheng et al., 8 May 2026, Mai et al., 1 Dec 2025).

1. Conceptual motivation

The central motivation is that advanced agentic intelligence depends on interaction with diverse environments rather than on static instruction following. One formulation states that diverse real-world APIs demand precise, robust function-calling intelligence, and that the breadth of function-calling competence is closely tied to the diversity of environments in which agents are trained. Another formulation identifies “task scarcity” in sandbox settings where no predefined task or reward exists, making task generation itself a prerequisite for reinforcement learning. In autonomous scientific discovery, a further shift is proposed: the primary bottleneck is no longer prescribing the agent’s internal workflow, but engineering the surrounding environment—its resources, interfaces, and constraints—so that productive behaviors are amplified and harmful behaviors are suppressed (Fang et al., 16 Sep 2025, Mai et al., 1 Dec 2025, Xin et al., 11 Jun 2026).

Within this literature, environment-task discovery is not limited to synthesizing natural-language prompts. It includes constructing executable tool suites, inferring latent dependency structure, defining verifiable end states, inducing proxy goal distributions, and exposing feedback channels that make long-horizon search tractable. This suggests that the problem is simultaneously one of data generation, systems design, and experimental control.

A further conceptual extension appears in scientific contexts. “Embodied science” argues that discovery is inherently physical and long-horizon, and that current computational approaches are misaligned when they reduce discovery to isolated prediction tasks. Its proposed closed-loop framing indicates that future environment-task discovery systems may have to couple perception, language, action, and discovery in laboratory settings rather than only in software simulators (Zhuang et al., 20 Mar 2026).

2. Formal representations of environments, tasks, and trajectories

A recurring abstraction is the agentic environment as an executable state-transition system. In GAIS, an environment is the tuple

$\mathcal E=(A,T,S),$

where $A$ is the set of executable tools, $T:S\times A\to S$ is a deterministic transition function, and $S$ is the latent state space; an agentic task is the triple

$\tau=(s_0,u,p),$

where $s_0\in S$ is the initial state, $u$ is the user intent, and $p$ is a domain policy, with a ground-truth trajectory $\tau^*$ given by a tool sequence achieving $u$ from $A$0 without violating $A$1. AgentScaler adopts a closely related but database-centric formulation: any function call is a read/write operator over an underlying database $A$2, each tool is assigned $A$3, and a task specifies an initial database state $A$4, a user intent $A$5, and a gold sequence of tool calls yielding a verifiable ending state $A$6 (Shi et al., 1 Jun 2026, Fang et al., 16 Sep 2025).

Trajectory structure is also formalized explicitly. AgentScaler defines an agentic trajectory, or experience, as an alternating user–assistant sequence

$A$7

with assistant turn

$A$8

where $A$9 are function-call tokens, $T:S\times A\to S$0 are tool-response tokens, and $T:S\times A\to S$1 is the final natural-language response. In CLI-Gym, the environment state is represented as

$T:S\times A\to S$2

where $T:S\times A\to S$3 is the base Docker image, $T:S\times A\to S$4 the Dockerfile, and $T:S\times A\to S$5 the installed codebase; the task-generation problem is then defined as inverting a healthy state into a failing state under agent control. In CuES, the environment is reduced to a task-free sandbox

$T:S\times A\to S$6

and task generation is formalized as learning a mapping

$T:S\times A\to S$7

that induces a proxy goal distribution over latent goals $T:S\times A\to S$8 when neither rewards nor named tasks are available (Lin et al., 11 Feb 2026, Mai et al., 1 Dec 2025).

Some works specialize the formulation further to make discovery tractable. AutoTTS casts width–depth test-time scaling as a deterministic, finite-horizon MDP or controller-synthesis problem with state

$T:S\times A\to S$9

where $S$0 is the question, $S$1 the number of branches, $S$2 the active branches, $S$3 their depths, and $S$4 the revealed probe signals; the action space includes $S$5, $S$6, $S$7, $S$8, and $S$9 under a cost budget $\tau=(s_0,u,p),$0 (Zheng et al., 8 May 2026).

These formalisms share three elements: an executable state, a constrained action interface, and a verifiable notion of success. Their differences lie in what is treated as latent—database state, runtime configuration, retrieval context, or proxy goals—and in how strongly the environment is grounded in external systems.

3. Environment construction and task induction mechanisms

The dominant methodological distinction is between unconstrained language-model generation and grounded environment construction. Several systems explicitly reject the former. GAIS argues that unconstrained synthesis often degenerates into biased random sampling of a model’s internal priors, failing to capture real-world diversity and long-horizon difficulty; its response is a two-phase grounding mechanism anchored in Model Context Protocol servers. AgentScaler begins from large API collections and materializes them into executable simulators. CLI-Gym begins from healthy containers and deliberately degrades them. CuES begins from structured environment descriptions and exploratory interaction traces. GSM-Agent constructs a controllable retrieval world from decomposed GSM8K premises, while NASA-EO-Bench derives research-query tasks from a knowledge graph linking publications, datasets, and tools (Shi et al., 1 Jun 2026, Fang et al., 16 Sep 2025, Lin et al., 11 Feb 2026, Zhu et al., 26 Sep 2025, Yu et al., 2 Jul 2026, Mai et al., 1 Dec 2025).

Environment source	Discovery mechanism	Output
Raw API inventories	Tool dependency graph, Louvain partitioning, schema materialization, random directed walks	$\tau=(s_0,u,p),$1 domains; hundreds of thousands of verifiable tasks
MCP server repositories	Protocol-anchored tool extraction, structure-guided planning, adversarial policy injection	$\tau=(s_0,u,p),$2 environments; $\tau=(s_0,u,p),$3 tasks
Healthy Docker environments	Agentic environment inversion with execution feedback and failing tests	1,655 CLI tasks
GSM8K problems	Premise sharding, document generation, retrieval database construction	$\tau=(s_0,u,p),$4 documents; 7,323 filtered problems
NASA EO-KG	Citation-grounded query–dataset extraction and abstract-to-query generation	47,654 positive pairs; 21,272 task-based queries
Task-free sandbox traces	Curiosity-driven exploration, task abstraction, re-execution, goal rewrite	Synthesized goal datasets for AppWorld, WebShop, and BFCL

AgentScaler’s pipeline is explicitly graph-based. It gathers approximately 30,000 real-world APIs, represents each tool’s parameter list as a vector, builds an undirected graph with cosine-similarity edges, runs Louvain community detection to obtain more than 1,000 tool communities, synthesizes a database schema for each domain, emits executable Python implementations for each tool, and then creates tasks by random directed walks on the tool-dependency graph. GAIS is grounded more tightly in real protocol implementations: it parses roughly 1,000 MCP repositories, rewrites tools into executable Python functions while preserving I/O schemas exactly, runs unit tests, scores tools on a 1–5 difficulty scale, filters out trivial or underspecified environments, constructs dependency graphs, and generates both compliant and adversarial tasks by depth-first walks from high-degree or high-difficulty anchor tools (Fang et al., 16 Sep 2025, Shi et al., 1 Jun 2026).

CLI-Gym introduces a distinct inversion paradigm. It treats a Dockerfile as a linear action sequence that transforms a poor environment into a healthy one, and then reverses that logic: starting from a gold state, an agent applies perturbations until a chosen subset of tests fails, records the Dockerfile patch, failed tests, and error messages, and packages the resulting broken state as a task. This converts environment corruption itself into an agentic task-generation process (Lin et al., 11 Feb 2026).

CuES is the most explicitly goal-inductive framework. It extracts a concept pool and principles from structured descriptions, performs curiosity-driven exploration with an environment memory tree, slides windows over traces to abstract multi-step task candidates, re-executes them with an Execution Agent and Judge Agent, and finally rewrites verified goals with incremental hints to create a small curriculum. GSM-Agent and NASA-EO-Bench, while aimed at benchmark construction rather than training-data synthesis in the narrow sense, show parallel patterns: both derive tasks from structured external sources, hide critical information from the initial prompt, and force the agent to recover it through controlled interaction (Mai et al., 1 Dec 2025, Zhu et al., 26 Sep 2025, Yu et al., 2 Jul 2026).

4. Verification, learning, and control

A defining feature of this area is that generated tasks are usually executable and verifiable, rather than merely plausible. AgentScaler filters trajectories by three criteria: validity control for well-formed dialogs, exact final-state agreement with the gold database state, and exact function-call sequence match for purely read-only tasks. Its supervised objective masks out human instructions and tool responses, optimizing only over tool-call and natural-language response tokens:

$\tau=(s_0,u,p),$5

followed by two-phase fine-tuning: cross-domain foundation learning and domain specialization (Fang et al., 16 Sep 2025).

GAIS embeds verification into environment construction itself. Tool conversion preserves I/O schemas exactly, unit tests are run, tools are discarded after repeated failure, and only environments with at least three actions and at least one tool of difficulty at least 3 are retained. It then enforces grounding constraints over synthesized trajectories so that dependency relations in the graph are respected. CLI-Gym similarly depends on execution feedback: the inversion loop terminates only when at least one target test fails, and the task package includes both fail-to-pass and pass-to-pass tests to guard against over-sabotage. CuES applies the most explicit post hoc quality control: tasks survive only if re-execution succeeds and the Judge Agent assigns a perfect “reward = 1.0” (Shi et al., 1 Jun 2026, Lin et al., 11 Feb 2026, Mai et al., 1 Dec 2025).

In controller-synthesis settings, verification becomes offline replay. AutoTTS pre-collects reasoning trajectories and probe signals, then evaluates candidate controller code without repeated LLM calls. Its controller is scored by an accuracy–cost objective

$\tau=(s_0,u,p),$6

and its search space is regularized by collapsing internal thresholds into a monotonic scalar $\tau=(s_0,u,p),$7. Full execution traces are preserved so that the proposing agent can diagnose whether it branched too often, pruned too early, or misused probe signals (Zheng et al., 8 May 2026).

GSM-Agent shows that the feedback channel itself can be redesigned. Its additional tools—Thinking, Explore, and Revisit—are not new external knowledge sources, but control instruments that bias the agent toward different traversal patterns over a document environment. The resulting perspective is that environment-task discovery often includes discovering the right interaction primitives rather than only the right task instances (Zhu et al., 26 Sep 2025).

5. Empirical regimes and application domains

Empirical evaluation spans function calling, CLI repair, retrieval, reinforcement learning, mathematical reasoning, and autonomous science. Reported metrics include pass@1, pass@3, accuracy on multi-turn simulations, Recall@$\tau=(s_0,u,p),$8, MRR, avg@8, greedy success rate, medal rate, and verified objective values. Across these domains, the measured gains are typically attributed not just to more synthetic data, but to better-grounded environments, stronger verification, or more informative control signals (Fang et al., 16 Sep 2025, Shi et al., 1 Jun 2026, Lin et al., 11 Feb 2026, Zheng et al., 8 May 2026, Zhu et al., 26 Sep 2025, Yu et al., 2 Jul 2026, Mai et al., 1 Dec 2025, Xin et al., 11 Jun 2026).

System	Setting	Headline result
AgentScaler	$\tau=(s_0,u,p),$9-Bench, $s_0\in S$0-Bench, ACEBench-en	AgentScaler-30B-A3B: 70.4 / 54.0 on $s_0\in S$1-Bench Retail / Airline; 70.2 / 60.0 / 55.3 on $s_0\in S$2 Retail / Airline / Telecom; 75.7 ACEBench-en Overall
GAIS	BFCL-V3, $s_0\in S$3-Bench, ACEBench-en	On Qwen3-8B-Base, 37.0 BFCL-Base, 46.5 Retail, 53.2 Normal, 7.5 Agent; with only 7K synthesized samples
LiberCoder	Terminal-Bench 1.0 / 2.0	Qwen3-32B: 10.3% $s_0\in S$4 38.9%; Qwen3-235B: 25.0% $s_0\in S$5 46.1%
AutoTTS	Held-out math reasoning	Improved held-out accuracy–cost trade-off in 7 of 8 Qwen3 settings; discovery cost $39.9 and 160 minutes
GSM-Agent	Controllable search benchmark	Frontier models around 67% accuracy; tool-augmented setups add 5–10 pp
NASA EO agentic search	NASA-EO-Bench and reranking subset	Hybrid retrieval: R@10 = 0.4275, MRR = 0.2918; agentic rerank: MRR $s_0\in S$6 on a stratified $s_0\in S$7 subset</td> </tr> <tr> <td>CuES</td> <td>AppWorld, WebShop, BFCL v3</td> <td>Average success: avg@8 50.70, greedy 51.16 vs. 34.50 and 35.01 on original tasks</td> </tr> <tr> <td><a href="https://www.emergentmind.com/topics/eurekagent" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">EurekAgent</a></td> <td>Autonomous <a href="https://www.emergentmind.com/topics/scientific-discovery" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">scientific discovery</a></td> <td>26-circle packing score 2.635999 at \$10.46 API cost; MLE-Bench Lite any-medal 85.71%, gold 71.43%

Several cross-domain regularities appear. AgentScaler reports that domain specialization adds approximately 5 percentage points on the ACEBench-en Agent subset and approximately 3 points overall, that out-of-distribution transfer to ACEBench-zh improves Overall from 74.2 to 81.5, and that long-horizon difficulty manifests as an approximately linear accuracy drop as the number of tool calls increases. GAIS reports superior data efficiency relative to Nemotron, matching a BFCL-V3 score of 21.8 with 1K samples where Nemotron needs 3.4K, and matching 24.0 with 2K where Nemotron needs 4.8K. CLI-Gym reports that environment diversity matters more than sheer trajectory count under a fixed 100-trajectory budget. GSM-Agent reports a strong positive correlation between revisit ratio and accuracy, $s_0\in S$8, and a correlation of approximately $s_0\in S$9 between changes in revisit ratio and changes in accuracy under tool-augmented scaling (Fang et al., 16 Sep 2025, Shi et al., 1 Jun 2026, Lin et al., 11 Feb 2026, Zhu et al., 26 Sep 2025).

The domain range is itself significant. NASA’s deployed geoscience system shows that the same environment-task logic can govern dataset discovery, where a neural scorer, BM25, and zero-shot agentic reranking are layered over a knowledge-graph-derived benchmark. EurekAgent shows that environment engineering can support metric-driven autonomous discovery in mathematics, GPU kernel engineering, and machine learning engineering, with hidden evaluators, budget controls, persistent artifacts, and human supervision infrastructure (Yu et al., 2 Jul 2026, Xin et al., 11 Jun 2026).

6. Limitations, misconceptions, and research directions

A common misconception is that more interaction, or more synthetic data, is sufficient by itself. The literature is more specific. GAIS argues that unconstrained LLM synthesis drifts toward low-difficulty or biased constructions unless it is anchored in external protocols. GSM-Agent reports only weak correlation between exploration and accuracy, but a strong positive correlation between revisit and accuracy and a negative correlation between exploitation and accuracy. These results indicate that the structure of interaction matters at least as much as its volume (Shi et al., 1 Jun 2026, Zhu et al., 26 Sep 2025).

Another persistent issue is the fidelity–executability trade-off. GAIS notes that code conversion may introduce subtle bugs in domain semantics, even though unit tests and interface checks preserved at least 94% core functionality. AgentScaler’s environments are offline simulators, and its authors explicitly note that real-world deployment through MCP servers and under rate limits adds further complexity. CLI-Gym similarly identifies remaining difficulty in scientific-computing and gaming CLI tasks, despite gains in dependency and configuration failures (Shi et al., 1 Jun 2026, Fang et al., 16 Sep 2025, Lin et al., 11 Feb 2026).

Several frameworks also expose scalability limits. CuES depends on a reliable structured environment description $u$0 and warns that its memory tree may become intractable in large or continuous state spaces. AutoTTS depends on pre-collected trajectories and probe signals, which makes discovery cheap but ties the method to a replayable search environment. These constraints suggest that some forms of environment-task discovery are best viewed as environment compression: they make search tractable by restricting what can vary and what can be replayed (Mai et al., 1 Dec 2025, Zheng et al., 8 May 2026).

Safety and oversight are not peripheral concerns. EurekAgent treats reward hacking, evaluation tampering, same-round copying, and unbounded resource consumption as environment-design failures. Its response is permissions engineering, isolated evaluation through a one-way grade(candidate) API, filesystem and Git-based artifact engineering, explicit API-cost and time budgets, and human-in-the-loop interfaces for intervention and auditability. In that framing, the environment is not only a source of tasks but also a mechanism for shaping acceptable research behavior (Xin et al., 11 Jun 2026).

The forward-looking agenda is correspondingly broad. AgentScaler identifies reinforcement learning and larger model scales as open directions. GAIS proposes expansion to streaming and real-time environments and schema extraction beyond MCP. CLI-Gym proposes extension to GUIs and cloud APIs, richer multimodal feedback, and reinforcement learning over degradation paths. CuES identifies on-policy or continual re-synthesis as a natural next step. “Embodied science” implies a further extension from software tasks to physical discovery loops in the life and chemical sciences, where empirical validation becomes the environment’s final grounding signal (Fang et al., 16 Sep 2025, Shi et al., 1 Jun 2026, Lin et al., 11 Feb 2026, Mai et al., 1 Dec 2025, Zhuang et al., 20 Mar 2026).