Tool Discovery Agents

• Tool Discovery Agents are autonomous systems that dynamically search, select, and compose external tools using reinforcement learning and semantic matching.
• They integrate methods like decentralized registries, vector indexing, and schema matching to enable efficient tool retrieval and chaining.
• They adapt scalable workflows through recursive retrieval and composite synthesis, improving performance on benchmarks and dynamic task demands.

Tool Discovery Agents are autonomous or semi-autonomous artificial agents equipped with mechanisms for identifying, selecting, composing, and often inventing or enhancing external computational tools, APIs, or other agents, given dynamic task demands and open-ended tool libraries. These agents comprise a central research area in automated reasoning, multi-agent systems, reinforcement learning, LLM orchestration, and the Internet of Agents (IoA). Tool discovery is distinct from mere tool use: it involves the ability to autonomously search, retrieve, select, chain, and potentially invent new tools or agent capabilities, often under uncertainty and at scale.

1. Foundations: Concepts and Formal Models

Tool discovery is formally distinguished from tool innovation and static tool-use by an agent’s ability to search, infer, and dynamically compose tools as part of solving new or evolving tasks. Under the Active Inference (AIF) formalism, tool discovery is the process by which an agent “happens upon” a useful external object or action sequence through active exploration, minimizing expected free energy (G) that trades off epistemic value (information gain) and pragmatic value (utility). The generative model’s state-transition parameters φ are learned during exploration, corresponding to Bayesian parameter inference, until the conditional distribution $P(s_{t+1}|s_t,a_t;φ)$ accurately supports reliable policy selection. By contrast, tool innovation arises from reconfiguring the latent factorization of the agent’s internal model to support "offline" composition and generalization to never-before-seen affordances, and thus, novel tool creation (Collis et al., 2023).

In multi-agent architectures, capability discovery is formalized as a two-stage process: (1) autonomous capability announcement, where each agent registers a structured, machine-interpretable description of its skills (functional APIs), roles, and non-functional constraints; (2) task-driven capability discovery, involving context-aware search, ranking, and composition to identify agents or tools matching a presented task intent. The matching process is governed by a utility function incorporating semantic relevance, compositional coverage, operational feasibility, and cost constraints (Guo et al., 24 Nov 2025).

2. Core Mechanisms and Architectural Patterns

Tool Discovery Agents operate in highly diverse computational environments, but exhibit structural parallels across domains:

• Registry and Metadata Layer: A decentralized or centralized registry houses tools/agent skill descriptions, typically structured as JSON/ontology entries (function names, input/output schemas, domain, version). Agents interact with this registry at runtime, dynamically querying, registering, or updating available tools (Wang et al., 15 Mar 2026, Guo et al., 24 Nov 2025).
• Retrieval and Ranking: Tool selection uses dense embedding-based similarity (cosine or ℓ₂), often following a hierarchical (coarse-to-fine) or vector-store retrieval pipeline. For instance, an agent first selects relevant servers, then ranks tools within those servers according to task alignment (Fei et al., 1 Jun 2025, Ocker et al., 2024). Product-quantization and inverted-file indexing enable fast vector retrieval at Internet scale (Guo et al., 24 Nov 2025).
• Semantic and Schema Matching: Many frameworks rely on input/output schema compatibility, either through explicit param-key overlap or semantic embedding alignment. Multi-parent synthesis and schema-overlap mechanisms enable cross-workflow and cross-agent tool chain composition (Wang et al., 15 Mar 2026).
• Iterative/Recursive Tool Discovery: Rather than loading all tool schemas into prompt context (costly and infeasible for large |T|), agents issue tool search queries recursively or proactively, and only retrieve and inject top-k relevant tool schemas as needed. This minimizes prompt size and supports adaptation to large tool libraries (Fei et al., 1 Jun 2025, Ocker et al., 2024, Li et al., 24 Oct 2025).
• Autonomous Tool Creation and Mutation: Systems such as Tulip Agent provide Create, Read, Update, and Delete (CRUD) interfaces, enabling agents to dynamically add, modify, or retire tools at runtime, supporting adaptive evolution of toolsets (Ocker et al., 2024). Composite tool synthesis by mining effective action routines is similarly employed for experience-driven expansion (Fan et al., 6 Mar 2026).
• Central Coordination vs. Decentralized Emergence: Tool discovery may be orchestrated by a governed hierarchy, as in Mozi’s supervisor-worker architecture for drug discovery pipelines (Cao et al., 4 Mar 2026), or may emerge in a fully decentralized ecosystem through artifact/need broadcasting, as in ScienceClaw + Infinite (Wang et al., 15 Mar 2026).

3. Discovery Workflow and Learning Algorithms

A typical tool discovery agent executes the following recurrent algorithmic loop (domain-specific variants exist):

• Task or Need Encoding: Encode the current subtask, observation, or artifact need into an embedding space shared with available tool or agent skill embeddings (Guo et al., 24 Nov 2025, Fei et al., 1 Jun 2025).
• Tool Retrieval: Query the registry or vector store for top-k most similar tool embeddings (or schema/param matches) (Ocker et al., 2024, Li et al., 24 Oct 2025).
• Selection and Invocation: Select tools above a similarity threshold (τ). If no tool is sufficient, recursively decompose the subtask or issue an explicit tool request, as in proactive frameworks (e.g., MCP-Zero’s <tool_assistant> block) (Fei et al., 1 Jun 2025).
• Composition and Chaining: For complex or multi-stage requirements, chain tools into pipelines/workflows according to output type and schema compatibility, or synthesize composite actions from observed frequent subsequences (Fan et al., 6 Mar 2026).
• Learning and Credit Assignment: Leverage reinforcement learning or variational free energy minimization to assign credit to successful tool selections and orchestrate memory folding or experience replay for persistent adaptation (Li et al., 24 Oct 2025, Fan et al., 6 Mar 2026). Specialized RL objectives (e.g., ToolPO, GRPO) attribute fine-grained signal to tool-invocation tokens or composite usage.

4. Scalability, Indexing, and Internet-scale Discovery

At Internet scale (billions of agents), tool discovery demands rigorous indexing and secure, performant registry protocols:

• Vector Indexing: Scalable updatable vector indices (e.g., product quantization, inverted files) allow sub-linear query latency and constant-memory slices, supporting discovery with N in the 10⁶–10⁹ range (Guo et al., 24 Nov 2025).
• Internet-native Protocols: The Domain Name System (DNS) can act as an authoritative substrate for agent and tool metadata discovery. By binding tool capabilities, endpoint addresses, protocol versions, and authentication anchors (DNSSEC, DANE) to FQDN records, the entire discoverable tool profile of an agent can be retrieved in a single DNS UDP transaction, typically ~1 ms, covering up to ~1 000 B of capability metadata per agent (Seethiraju et al., 1 Jun 2026).
• Provenance and Lineage Auditing: In decentralized agent societies, lineage is recorded as a global DAG (artifact→parent) with conflict, redundancy, and stagnation resolution (mutation layer), supporting traceability, convergence, and workflow deduplication (Wang et al., 15 Mar 2026).
• Security, Privacy, and Economics: Discovery agents must integrate verifiable credentials, privacy-preserving queries (e.g., zero-knowledge proofs), and reputation or incentive mechanisms to facilitate trust and healthy collaboration at scale (Guo et al., 24 Nov 2025, Seethiraju et al., 1 Jun 2026).

5. Evaluation, Performance, and Empirical Results

Tool discovery agent frameworks are evaluated on criteria including discovery accuracy, latency, prompt/token efficiency, orchestration accuracy, reproducibility, and generalization:

• Accuracy and Token Cost: On benchmarks such as API-Bank and ToolBench, agent frameworks employing retrieval-based open-set tool discovery consistently outperform baseline systems that inject all tool schemas into the prompt, typically reducing token consumption by >98% and maintaining or improving accuracy (>95% top-1 selection in MCP-Zero, DeepAgent, and Tulip Agent evaluations) (Fei et al., 1 Jun 2025, Li et al., 24 Oct 2025, Ocker et al., 2024).
• Autonomous Tool Synthesis: Experience-driven composite tool discovery (MACRO) yields significant improvements in multi-step orchestration accuracy (e.g., +2–6 percentage points balanced accuracy vs. static toolchain baselines) and cross-domain generalization (Fan et al., 6 Mar 2026).
• Provenance and Reproducibility: Architectures with built-in logging, deterministic replay, and strict data contracts (Mozi) support robust, auditable workflows in high-stakes domains such as drug discovery (Cao et al., 4 Mar 2026).
• Scalability and Latency: DNS-based discovery mechanisms exhibit lookup latency of ~5 ms and reliably encode full capability metadata for >99.7% of agents within the wire-size limits of a single UDP transaction (Seethiraju et al., 1 Jun 2026). Vector-based registries enable average query times ≤8 ms for 1,000–4,000 agents (Guo et al., 24 Nov 2025).

6. Open Problems and Future Directions

Despite rapid progress, key challenges and research questions persist:

• Semantic Matching and Compositionality: Existing systems frequently employ syntactic key overlap for schema compatibility. Embedding-based and learned semantic compatibility measures are under development (Wang et al., 15 Mar 2026).
• Autonomous Structure Learning: Extending tool discovery to support autonomous factorization, expansion, or pruning of model structure—enabling adaptation to new tool affordances or primitive sets—remains an active area (Collis et al., 2023).
• Cross-domain and Hierarchical Composition: Integrating agents from disparate domains (e.g., combining medical and robotics tools) using graph neural networks and meta-learning for one-shot orchestration is recognized as a critical path (Guo et al., 24 Nov 2025).
• Economic Incentives and Trust: To sustain healthy, open agent economies, future work focuses on token-based incentives, reputation protocols, and robust verification mechanisms (Guo et al., 24 Nov 2025).
• Integration with Existing Infrastructure: Adoption of standard Internet protocols (DNS), robust auditing, dynamic registry updates, and scalable storage/indexing is a focus for global Internet of Agents deployments (Seethiraju et al., 1 Jun 2026).

Tool discovery agents thus comprise a convergent line of research, blending theoretical formalism, RL-driven orchestration, scalable information retrieval, decentralized governance, and practical software architectures. Their evolution is central to the development of robust, adaptive, and scalable AI systems capable of dynamic composition and continual innovation in open-ended computational environments.