SAGE boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

SAGE is in scope as a builder or improvement plane: it trains a model through skill-augmented agent rollouts and ships the corresponding task-solving agent and data-generation paths. The intended analysis covers how reusable Python functions are produced, admitted, retained and consumed during training and evaluation, alongside how rollout rewards feed model updates. This is a whole-system boundary over the shipped SAGE repository, not merely its skill-extraction API. AppWorld callers, prompts, evaluators and SFT patches supplied in this repository are included even though installation overlays some of them onto external repositories (SRC-1, README.md, sage/run_sage_2nodes.sh, patches/appworld/experiments/code/skill_library_agent/agent.py).

The source declares its identity and training mechanism:

This repository contains the official implementation of SAGE (Skill Augmented GRPO for self-Evolution), introduced in our paper Reinforcement Learning for Self-Improving Agent with Skill Library. --- README.md

The shipped RL launcher calls verl.trainer.main_ppo; that entry point constructs RayPPOTrainer, initializes its workers and calls fit. The trainer imports the AppWorld generation manager, feeds generated rewards into advantage computation and calls the actor update. This establishes model-dependent operation, rather than a passive function store (SRC-1, sage/run_sage_2nodes.sh, sage/verl/trainer/main_ppo.py, sage/verl/trainer/ppo/ray_trainer.py).

actor_output = self.actor_rollout_wg.update_actor(batch) --- sage/verl/trainer/ppo/ray_trainer.py

The training loop and evaluation loop are both included, but their retention mechanisms must not be conflated. Training initializes skill dictionaries for a rollout and passes them across two subtask iterations. The batch environment extracts functions from non-failing execution output on the first subtask, then executes and supplies retained functions in the second subtask's prompt. Evaluation instead keeps an agent-level function list, appends function records to JSONL, optionally saves embeddings, and can load a separately located predefined library on construction. The analysis must distinguish in-process reuse, output persistence and predefined-library loading; these paths alone do not establish automatic restart recovery from evaluation output (SRC-1, sage/llm_agent/appworld_generation.py, sage/llm_agent/app_world/app_world_env.py, patches/appworld/experiments/code/skill_library_agent/skill_library_agent.py).

skill_libraries = [{} for _ in range(len(gen_batch["task_ids"])//2)] skill_counts = [0 for _ in range(len(gen_batch["task_ids"])//2)] for subtask_iter in range(2): --- sage/llm_agent/appworld_generation.py

if os.path.exists("./predefined_skill_library/skill_library_functions.jsonl"): --- patches/appworld/experiments/code/skill_library_agent/skill_library_agent.py

The evaluation config selects the skill-library agent. Its shipped runner constructs one agent and calls solve_tasks; the base agent executes generated code, evaluates tasks and invokes skill extraction when configured. Expert generation adds scenario rollouts, verification and dataset extraction, with SFT configuration consuming the resulting dataset. These are included state producers and consumers, not excluded applications (SRC-1, patches/appworld/experiments/configs/sage_test_normal.jsonnet, patches/appworld/experiments/code/skill_library_agent/run.py, patches/appworld/experiments/code/skill_library_agent/agent.py, patches/appworld/extract_expert_dataset.py, patches/LLaMA-Factory/examples/train_full/qwen2_5_32B_appworld.yaml).

Prompts are part of the mechanism, not independent evidence that every requested property is enforced:

(3) If no suitable function exists in the skill library, you should define a new function first and then call it to complete the step. --- patches/appworld/experiments/prompts/skill_library_agent.txt

Shipped-path coverage

The tracked tree was enumerated at the frozen commit. Included directories below remain within the evidence boundary in full; named paths document the material responsibilities initially traced, not an allowlist. “Not yet inspected” is not an access gap.

Dependencies and evidence limits

The frozen revision is 3c9244e82244abb1adc5467ee601a03ba0f433a0; the applicability cutoff is 2026-10-15, taken from the supplied run date, not a claim about when this commit was authored. The acquired checkout's HEAD matches the pin and its working tree reports no changes. Overall evidence is code-grounded, with README and prompts supplying doctrine/design. No target execution was performed; no observed-run or causal-experiment layer is established.

Upstream AppWorld and LLaMA-Factory implementations are external dependencies, not additional frozen sources. The README identifies upstream checkout revisions, but those references do not register their contents here. AppWorld supplies the environment, task data and underlying evaluator; SAGE supplies the inspected orchestration and patches. Missing upstream implementation evidence prevents conclusions about sandbox isolation, benchmark-test adequacy and full patched-install compatibility. SFT's upstream training machinery is likewise not attributable to SAGE's patches (SRC-1, README.md, patches/appworld/src/appworld/cli.py, patches/LLaMA-Factory/run_sft.sh).

Model weights, downloaded AppWorld data, predefined skill-library files, hosted model responses and generated experiment outputs are not frozen in this source set. Their absence prevents conclusions about actual learned competence, library quality, empirical reuse benefits, achieved evaluation scores or reproducibility on particular hardware. The shipped code is evidence for what the implementation attempts, not for successful training or causal improvement. Ray, PyTorch, vLLM, model-provider adapters and retrieval models are runtime dependencies; only their shipped integration code falls within SRC-1. The linked paper and linked external repositories have not been acquired and supply no separate evidence in this run (SRC-1, README.md, sage/run_sage_2nodes.sh, patches/appworld/experiments/code/skill_library_agent/skill_library_agent.py).

Source register