INQUIRING LINE

Code is organized by file and function, but any one behavior hides across many files — what if repos were built around behaviors instead?

How can we reorganize repositories to make behaviors easier to locate?

This explores how the way code is organized — around structure vs. around behavior — determines whether you can find all the places a single behavior lives, and what the corpus says about reorganizing toward the latter.


This is really a question about a mismatch: code is stored by structure (files, functions, stages) but we think about it by behavior ("how does this system handle retries?"). The corpus names this directly — the hard part of changing a system isn't writing the edit, it's finding every scattered place a behavior is implemented, because a single behavior gets smeared across many files and stages Why is finding distributed behavior code so hard?. Reorganizing the repository, then, means building a representation that maps behaviors to their code locations instead of leaving that map implicit in the folder tree.

The strongest concrete result: when a harness is reorganized around a behavior-to-code mapping, even a weaker planning model matches a much stronger one at locating the right code — win rates jumped 10–19 points while the planner read fewer tokens Can explicit behavior maps help weaker planners compete with stronger models?. The lesson is that locatability is a property you can design into the repository's structure, not just a skill you hope a smart-enough model brings. Make the behavior boundaries explicit and the search problem mostly disappears.

Laterally, this is the same idea that shows up wherever agents accumulate reusable capability. Skill libraries store executable behaviors in an embedding-indexed catalogue and compose complex ones from simpler ones, which is what lets an agent keep learning without overwriting old skills Can agents learn new skills without forgetting old ones?. Agent Workflow Memory goes finer, extracting sub-task routines rather than whole tasks and compounding them hierarchically Can agents learn reusable sub-task routines from past experience?. In both, the win comes from the same move as behavior-centric harness organization: carve the repository at behavioral joints so units can be found, reused, and recombined.

A subtler point from the corpus is that a repository isn't best organized once and frozen — it can be actively curated toward locatability. SkillOS separates a trainable curator from the frozen executor, and that curator learns to reshape the skill store away from verbose generic entries toward crisp, actionable, cross-task logic Can a separate trained curator improve skill libraries better than frozen agents?. Reorganization becomes a learned, ongoing process, not a one-time refactor.

Worth knowing you didn't ask for: there may be no single best organization at all. StructRAG shows that different queries want different knowledge structures — tables, graphs, catalogues, plain chunks — and a router trained to pick the right structure per query beats any uniform layout, grounded in cognitive-fit theory Can routing queries to task-matched structures improve RAG reasoning?. And the flip side is a warning: retrieval over any repository fails structurally, not incrementally, because embeddings measure association rather than task relevance and have hard representational limits Where do retrieval systems fail and why?. So "make behaviors easier to locate" may mean maintaining several organizations and routing between them — not searching harder within one.


Sources 7 notes

Why is finding distributed behavior code so hard?

The core difficulty in evolving production harnesses is not generating edits but finding every code location that implements a behavior. Harnesses distribute single behaviors across files, functions, and stages, creating a representational mismatch between behavioral requests and structural code organization.

Can explicit behavior maps help weaker planners compete with stronger models?

A behavior-to-code mapping representation improved win rates by 10–19 points while reducing planner tokens by 8–13%. Weaker planners using this mapping matched stronger models' code localization across all precision and recall metrics.

Can agents learn new skills without forgetting old ones?

VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.

Can agents learn reusable sub-task routines from past experience?

Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.

Can a separate trained curator improve skill libraries better than frozen agents?

SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.

Show all 7 sources
Can routing queries to task-matched structures improve RAG reasoning?

StructRAG demonstrates that selecting knowledge structure type based on query demands—via DPO-trained router choosing among tables, graphs, algorithms, catalogues, and chunks—improves knowledge-intensive reasoning over standard retrieval. The approach grounds this in cognitive load and cognitive fit theory from cognitive science.

Where do retrieval systems fail and why?

RAG systems fail at three structural levels: adaptive triggering (fixed intervals waste context), semantic-task mismatch (embeddings measure association, not relevance), and mathematical limits (embedding dimension constrains representable document sets). These require fundamentally different retrieval approaches, not tuning.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a systems analyst investigating a still-open question: how should we reorganize repositories so that *behaviors* — not files — become easy to locate? Treat the claims below as dated and perishable, not current truth.

What a curated library found — and when (dated claims, not current truth), spanning ~2024–2026:
- The real bottleneck in changing a system is behavior localization: one behavior is smeared across many files and stages, so finding every site dominates the edit (~2026).
- Reorganizing a harness around an explicit behavior-to-code mapping let a *weaker* planner match a stronger one — win rates rose 10–19 points while reading fewer tokens (~2026).
- Skill libraries store executable behaviors in an embedding-indexed catalogue and compose complex ones from simple ones; Agent Workflow Memory extracts sub-task routines and compounds them hierarchically (~2024).
- SkillOS decouples a trainable curator from a frozen executor, learning to reshape the store toward crisp, cross-task logic — reorganization as ongoing process, not one-time refactor (~2026).
- StructRAG: different queries want different structures (tables, graphs, chunks); a per-query router beats any uniform layout, and plain retrieval fails structurally because embeddings measure association, not task relevance (~2024).

Anchor papers (verify; mind their dates): Agent Workflow Memory (2024); StructRAG (2024); From Model Scaling to System Scaling: Scaling the Harness (2026); SkillOS (2026).

Your task:
(1) Re-test each constraint. For every finding, judge whether newer models, training, tooling, orchestration (memory, caching, multi-agent), or evaluation has relaxed or overturned it; separate the durable question from the perishable limitation, cite what resolved it, and say where a constraint still holds.
(2) Surface the sharpest live disagreement — one-best-organization vs. maintain-many-and-route, and whether structural retrieval failure is fundamental or now patched — using contradicting work from the last ~6 months.
(3) Propose 2 research questions that assume the regime may have moved.

Cite arXiv IDs; flag anything you cannot ground in a real paper.