Agentic ML Task Execution

• Agentic ML task execution is defined as orchestrating autonomous agents into dynamic coalitions to execute complex ML workflows across distributed environments.
• It employs a decentralized, local-information algorithm to form coalitions that meet capability requirements while minimizing cost and ensuring scalability.
• The framework integrates economic models and reward allocation mechanisms to guarantee cost-effectiveness and individual rationality, promoting stable, incentive-aligned execution.

Agentic ML task execution refers to the design, orchestration, and evaluation of distributed ML workflows executed by autonomous, tool-using agents—often built atop LLMs—capable of planning, reasoning, inter-agent communication, and robust multi-step action across heterogeneous environments. This paradigm transcends static, single-model pipelines by embedding agents within cloud-, edge-, and network-native infrastructures that emphasize modularity, economic viability, interoperability, and principled coalition formation.

1. Network-Native Agentic Workflow Model

A foundational advance in the field is the formalization of the network-native model of agentic collaboration. Let $G=(\V,\E)$ be the undirected communication graph whose nodes each host a local agent pool $\A_i=\{a^i_1,\dots,a^i_{n_i}\}$ with per-agent capability sets $C(a)\subseteq\T$, where $\T$ is the global “capability space.” A task is specified at an initiator node $i_0$ by a multiset of required capabilities $\R_q$, mapping to subtask requirements in a directed acyclic workflow graph $W_q=(\T_q,\D_q)$.

A coalition $\C_q\subset\V$ executes the workflow if it jointly covers the capability requirements, adheres to a $k$-hop locality constraint (defining the feasibility radius $\rho_q$), and satisfies economic implementability via nonnegative effort, effort costs $c_i(u_i)$, communication costs $C_i^{\mathrm{comm}}$, and an incentive-compatible reward allocation $\{w_i^q\}$ that guarantees budget balance and individual rationality. This rigorous construct of workflow-coalition feasibility unifies coverage, locality, and economic constraints into a single operational envelope (Yang et al., 3 Feb 2026).

2. Decentralized Coalition Formation and Scalability

To efficiently select a minimum-effort, feasible agent coalition under the constraints above, the framework proposes a decentralized, local-information algorithm. For a task-initiating node, Algorithm 1 (“Workflow-Coalition Formation”) explores $k$-hop neighborhoods, enumerates candidate coalitions satisfying capability coverage, and, for each, verifies subtask agent assignments and feasibility under cost/reward models. Among all successful candidates, it selects the coalition minimizing total exerted effort. Empirically, the search space, though combinatorial in the worst case $O(2^{|\N^{(k)}|})$, is substantially reduced in sparse and capability-rich settings; convergence occurs within a handful of rounds, even as the network scales (Yang et al., 3 Feb 2026).

This decentralized design means that nodes require only local capability, cost, and bid summaries—no global view of the graph—drastically enhancing scalability, robustness to failures, and specialization at cloud and edge nodes.

3. Economic Incentives and Feasibility

Economic implementability is formalized by assigning effort levels $u_a \geq 0$ to agents, incurring costs at their respective nodes. The workflow reward $R_q = r_q(O_q)$ is distributed via allocations $\{w_i^q\}$ that strictly satisfy: $\sum_{i\in\C_q} w_i^q = R_q,\quad w_i^q \geq c_i(u_i) + C_i^{\mathrm{comm}},\; \forall i \in \C_q$ where output quality is a function of aggregate agent efforts. The positive surplus constraint enforces that all participants—after accounting for costs—are at least as well off as non-participants.

This incentive-aligned mechanism is essential for stable, self-organizing agentic execution in open networks where agents may have autonomy regarding coalition participation.

4. Interoperability and Integration (C+MCP Layer)

The coalition-formation layer operates strictly above the Model Context Protocol (MCP), which standardizes tool discovery and invocation. By introducing abstract capabilities, cost/quality parameters, and economic constraints into the MCP tool directory, the agentic coordination layer selects not only “which tools to invoke” but also “in what order and with what coalition—subject to reward/cost feasibility.” MCP remains unchanged at the call/invocation level, ensuring seamless composability across organizational boundaries and technical platforms (Yang et al., 3 Feb 2026).

5. Empirical Case Study: Healthcare Workflow

The framework was benchmarked on an Erdős–Rényi network of $N=40$ nodes with randomized assignment of the capability set $$\{\mathrm{OCR},\mathrm{RAD},\mathrm{DX},\mathrm{VAL},\mathrm{CONS}\}$$, representing sequential stages of a radiology diagnosis workflow. Economic parameters—agent deliberation effectiveness, quadratic latency/cost models, and reward curvature—enabled realistic coalition selection. Key findings:

• Feasibility could be achieved with $\C_q^* = \{0, 20, 22\}$ for $k = 2$, achieving total cost ≈ 4.51, reward ≈ 8.25, surplus > 0.
• As per-node capability breadth increased, both coalition size and required hop-radius decreased, illustrating improved scalability.
• In Monte Carlo sweeps, coalition search converged rapidly to near-optimal cost.

These results demonstrate that agentic ML workflows, with distributed coalition formation, remain performant, scalable, and economically viable across cloud and edge topologies, with resilience to network sparsity (Yang et al., 3 Feb 2026).

6. Workflow Semantics, Task Assignment, and Execution

After coalition formation, the system executes a directed-acyclic workflow graph $W_q$ where each subtask $\tau_j$ is assigned to a specific agent by a function $\varphi$, requiring that $c(\tau_j) \in C(\varphi(\tau_j))$. Each agent executes local computations $O_a = f_a(I_a, u_a)$, corresponding to its assigned subtask, aggregates outputs upstream, and participates in reward sharing according to the coalition contract. The framework formally guarantees that task assignment is valid and executable, as all feasibility, assignment, capability, and economic constraints have already been enforced prior to execution.

Once the coalition has been instantiated, all subsequent tool invocations and data exchanges occur strictly within the prescribed subgraph, leveraging MCP for standardized calls (Yang et al., 3 Feb 2026).

7. Impact and Future Directions

The Internet of Agentic AI framework embeds workflow planning, incentive alignment, and network locality directly into the foundational structure of agentic ML task execution, with the following implications:

• It provides a scalable, decentralized alternative to monolithic, centrally planned agent architectures.
• It guarantees executional feasibility, reward- and cost-alignment, and resilience against fragmentation in multi-organization clouds.
• The rigorous, multi-factor feasibility construct sets a foundation for future extensions, including more complex economic objectives, adversarial agent behaviors, capability learning, and dynamic coalition reconfiguration.

By layering on top of protocols such as MCP and integrating incentive-compatible coalition formation with local decision-making, agentic ML task execution is poised for robust, interoperable deployment across diverse, real-world distributed infrastructures (Yang et al., 3 Feb 2026).