Écrire un commentaire

The authors argue that a skill package should sometimes run as a subagent in its own context window instead of being loaded into the main agent's context, and that this pays off only when the package carries procedural instructions with explicit input and output contracts (Sections 3.1 and 3.2). On a 64-task subset of SkillsBench, subagents beat inline agent skills with contract-based packages, lose with the curated ones, and degrade more slowly as distracting skills are added (Section 4, Figures 2 and 3).

I read it as a practitioner who runs coding agents every day, and my notes are about where the paper meets things I have written down or measured myself.

1. Context isolation is the right reason to spawn a subagent, and the paper says so plainly. Section 3.1 notes that Claude Code and Codex tend to use subagents for parallel work, not for keeping information out of the main window. My own working rule, in my practice guides, runs the other way: quality drops as the context fills, so heavy reading goes to a subagent and the main conversation stays clean. Figure 3 puts that habit under a controlled test, and it holds. One caveat, which is mine and not the authors': my rule came from practice, not from an experiment, so I cannot say how far Figure 3 generalises beyond SkillsBench.

2. The token overhead in Figure 4 (right) deserves a cost column, not only a token count. I measured 722 of my own agent sessions (150,902 model calls, 34.6 billion tokens; technical report, not peer-reviewed, DOI 10.5281/zenodo.22759216). More than 85% of the modelled cost there was context handling, meaning cache reads plus cache writes, not generation. In a separate set of 168 sessions, 94.4% of paid tokens were re-reads of context already sent (DOI 10.5281/zenodo.22712985). That second number is a share of tokens, not of money, because re-reads are the cheapest tokens. So a subagent that duplicates information into a fresh window and a main agent that re-reads a long window on every turn can cost very different amounts for the same token total. Splitting Figure 4 by fresh input, cache writes, cache reads and output, priced at one public price list, would show whether the overhead is real spend.

Fragmenting too far also costs. In a small controlled experiment (36 runs, DOI 10.5281/zenodo.22759217) I found a U-shaped curve: clearing context after every task came out roughly a third more expensive than clearing every three tasks, with 3 to 6 tasks per session on a plateau. The comparison point was chosen after the runs, and the agent wrote more tests under per-task clearing, so part of that gap may be extra work. Still, the pattern suggests the hybrid mode in Appendix A.1 could have an optimum in how many subagent calls a task gets. Figure 7 reports more skill calls per task under subagents; accuracy against the number of subagent calls would be worth a plot.

3. The output contract is also the obvious place for a check. Under Equation 7 the main agent sees only the subagent's last response, and its internal trajectory stays hidden (Sections 2 and 3.1). In my guides I keep two rules for this. Every task needs a check the agent can actually run, otherwise "looks done" is the only signal it has. And a model sounds equally sure when it is right and when it is wrong. A subagent that returns something outside its q_out will read exactly like one that did the job. Did the authors count how many subagent-mode failures came from a return that broke its output contract and was accepted anyway? A second fresh agent asked to refute the return, or a plain validator of q_out, is cheap next to a failed long-horizon task.

4. Short, concrete instructions. The curated packages that lose in Figure 2 describe relevant knowledge but rarely state inputs or outputs (Section 4). This matches what I see with instruction files: concrete rules are followed more reliably than vague ones, and rarely needed material belongs in a skill, not in the permanent file.

Competing interests

Yes: the text cites the author's own technical reports (DOI 10.5281/zenodo.22759216, 10.5281/zenodo.22759217, 10.5281/zenodo.22712985); no connection to the article's authors.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.