Study finds instructions account for 60.5% of coding-agent reading

• Instruction files and working notes accounted for 60.5% of the study's documentation interactions across 557 coding-agent sessions, as dair_ai reported.
• Documentation consultation was mostly self-initiated, 70.2% of cases, while only 7.5% followed a failure, according to dair_ai's study summary.
• Backpass turns repository session logs into a reviewable proposal for a project AGENTS.md, as kunchenguid's announcement describes.
• Installed skills impose a standing cost of 50 to 280 prompt tokens each, while rohanpaul_ai's paper summary says the proposed @skills protocol reserves permanent residency for automatic triggering.

API references accounted for just 1.3% of the interactions in the documentation study, while code was touched first 4.7 times as often in multi-commit pull requests that changed both code and docs. Kun Chen's Backpass field note treats transcripts as the training data for project instructions. The @skills paper proposes :save as a way to put an external skill into a Git-tracked project tree without making it permanent prompt context.

The paper defines the 60.5% figure over 3,033 documentation interactions extracted from 94,813 development events. Its second dataset classifies 690,260 file-level changes from 33,097 agentic pull requests.

The paper's classification of those 3,033 interactions is:

1. Agent-facing artifacts, including instruction files and working notes: 60.5%.
2. Classical technical documentation: 10.6%.
3. API references: 1.3%.

The study records 70.2% of consultation as self-initiated and 7.5% as failure-driven. In pull requests that changed both code and documentation, code was touched first 4.7 times more often.

Reading documentation had an adjusted odds ratio of 0.39 for immediate testing in the corpus, per dair_ai's summary.

Kun Chen built Backpass around a split between human-authored user preferences and a project file trained against what actually happened in that repository's sessions. The field note gives the project file a size budget and treats individual instructions as units whose usefulness can be checked against transcripts.

Backpass's announced loop is:

1. Sample prior agent sessions for a repository.
2. Distill the sessions' key learnings and losses.
3. Synthesize a proposed "gradient descent step" for project instructions.
4. Let a reviewer apply the change.

Chen said he maintains AGENTS.md and points CLAUDE.md to it in kunchenguid's reply. A later kunchenguid follow-up said Hermes transcript support was being explored.

The @skills paper separates content, persistence, and automatic triggering, arguing that only the last requires permanent prompt residency. It estimates fewer than 100 skill descriptions can trigger reliably, against 56,804 published skills.

At the authors' stated range of 50 to 280 tokens, ten resident skill descriptions consume 500 to 2,800 prompt tokens on every request. The proposed protocol offers three paths:

• A path reference reads a skill, subtree, or collection when it is needed.
• :save vendors a copy into the project's Git-tracked tree.
• :install is the resident, automatic-triggering path, associated in the paper with a .gitignore entry.

Instructions are one component of an agent harness. A harness analysis from ZhihuFrontier separates coupling into three layers:

• Format: tool-call protocols and schemas.
• Context structure: history layout, result truncation, and compaction.
• Control flow: planning, reflection, and stopping behavior.

Armin Ronacher's report on Pi tool calls found newer Claude models sometimes added invented fields to Pi's nested edits[] schema. The edit could be semantically correct while the call failed validation.

Two August papers move post-training into the deployment harness rather than reproducing its loop inside an RL system:

• Agent Lightning v1.0 places a proxy between the harness and model, recording the harness's calls while leaving the harness in charge of environment interaction.
• ClawGym II runs OpenClaw or Claude Code unchanged in sandboxes, intercepts serving-boundary calls, and reconstructs fragmented calls into trajectories for PPO or GRPO.

onusoz's benchmark proposal calls for a simple, stable common harness, naming mini-swe-agent, Terminus 2, and vanilla Pi as possible baselines. By contrast, steipete's reply argues that removing retained reasoning, compaction, or code-oriented tools can suppress behavior a model was trained to use.

A post from omarsar0 similarly calls model measurement through harnesses broken, while still favoring minimal Pi or Hermes Agent baselines. onusoz's multi-harness proposal offers a middle ground: several standard environments, such as a Bash-fluent and a code-mode-fluent class.

The onusoz follow-up graphic draws a line between a fixed-harness score, which measures a model through a common interface and resource set, and a complete vendor system's best achievable result. It says those questions need separate leaderboard columns.