Subagents vs Agent Skills: Executing Reusable Knowledge for Long-Horizon Agentic Tasks

The proposed alternative invokes skill packages through delegated agents instead of placing their instructions in the primary agent context.

Source: [1], [4]

The authors argue that modular interfaces and delegated execution can help constrain information flow as agent systems scale.

Source: [10]

Efficient communication between primary and delegated agents, along with the organization of growing skill libraries, remains an open design problem.

Source: [7]

The study asks how language-model agents can use reusable knowledge libraries effectively for extended tasks.

Source: [19]

A skill package combines a purpose description, procedural instructions, and supporting artifacts for discovery and task execution.

Source: [29]

Under direct skill execution, the instructions enter the primary context and the primary policy carries out the prescribed procedure.

Source: [11], [24]

Delegated execution initializes a distinct context from task input and skill instructions rather than exposing those instructions to the primary agent.

Source: [15], [22]

The delegated agent conducts its own tool-using reasoning and supplies only its final response to the primary agent.

Source: [23], [31]

When work can be divided into self-contained subtasks, isolated contexts can reduce the greatest context burden borne by any individual context.

Source: [25], [26]

This isolation raises overall token demand because relevant information must be supplied again across separate contexts.

Source: [20]

The proposed skill design specifies appropriate invocation, delegated procedure, and returned information through explicit interfaces.

Source: [17], [32]

Input and output contracts define a self-contained boundary governing what a delegated skill accepts and returns.

Source: [27]

Evaluation uses a benchmark of long-horizon agent tasks with curated human-authored skill packages.

Source: [13]

The reported evaluation focuses on the benchmark subset for which procedural packages with explicit interfaces were synthesized.

Source: [18]

The experiments use an agent harness to compare direct skill execution with delegated execution.

Source: [5], [9]

The evaluation measures task success, peak context demand, and total token cost across underlying language models.

Source: [5], [16]

For the curated packages without explicit input-output contracts, direct skill execution matched or exceeded delegated execution across the evaluated models.

Source: [21]

For synthesized procedural packages with explicit interfaces, delegated execution outperformed direct execution, with the strongest reported advantage among smaller models.

Source: [18], [21]

For stronger models under the synthesized-package setting, delegated execution reduced peak context demand on most tasks.

Source: [3]

Delegated execution consumed substantially greater total token cost because contexts require repeated task-relevant information.

Source: [14]

Among the tested library organizations, task-tree and hierarchical-graph arrangements performed best, while deeper hierarchy did not provide further observed gains.

Source: [6], [30]

Peak-context comparisons are less meaningful for weaker models when direct execution ends early and therefore has a short context without comparable task success; comparison is most appropriate where success is similar between execution modes.

Source: [28]

The performance comparison between synthesized and curated packages is uncontrolled because their content differs, so stronger synthesized-package results only suggest comparable or better package quality.

Source: [2]

As irrelevant tool options increased in the reported setting, delegated execution degraded more gracefully than direct execution.

Source: [8], [12]