docs.manual Guide

Baselines

The paper compares four learned-skill methods: AWM, ASI, SkillWeaver and WALT. Their integrations are not released here.

Stating that plainly is the honest thing to do. The cards below name the upstream repositories we started from. Commit pins, harness attachments, and the libraries as we ran them are not recorded here.

What is and is not released

Released
The four methods’ code as we ran themno
The four methods’ per-task results and tracesno (results-policy.md)
Their aggregate numbersyes — transcribed from the paper’s Table 1, labelled verified-paper
Upstream repositoryyes — below
MPCR, including all six ablation variantsyes, src/skillshift/agents/mpcr/
The frozen skill library ASI and MPCR shareyes, src/skillshift/agents/mpcr/actions/ and workflows/

Leaderboard entries transcribed from a paper carry the verified-paper label and no run receipt. They are not claims about this harness; they are the published numbers, reproduced verbatim, with their source named. See results-policy.md.

Method cards

AWM — Agent Workflow Memory

Skills are prompted textual workflows: induced task/action-trajectory pairs injected into the agent’s context. There is no executable invocation, so Average Skill Calls and Skill Failure Rate are undefined for AWM and Table 1 prints --.

ASI — Inducing Programmatic Skills for Agentic Tasks

Skills are executable Python functions, verified for correctness, usage and validity, then exposed as callable high-level actions alongside the primitives.

SkillWeaver

WALT

The fixed-library protocol

For the MPCR comparison, ASI first learns a library on src. The library is then frozen and shared: ASI uses its own selector, MPCR changes only test-time selection over the same skills and primitives. This isolates executor-side adaptation from additional skill acquisition, and it is why the comparison in the paper’s Table 3 is attributable to selection rather than to learning.

Reference agent

The paper has no no-skill baseline — all four methods carry a library. The built-in reference agent fills that gap: plain ReAct over the primitive action set only, ASC = 0 and SFR = null by construction. Its purpose is calibration, not comparison, and it does not appear on the leaderboard. See calibration.md for what skillshift calibrate does with it, and metrics.md for the metric definitions.