ACE boundary

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

Boundary and evidence

This run analyses how ACE produces, updates, persists, reloads and consumes a playbook of task-solving knowledge. The intended use is an implementation-grounded account of that adaptation loop, including its shipped evaluation callers, rather than an audit of benchmark performance. ACE qualifies as a memory/knowledge/context-engineering system: its generator consumes playbook text, its reflector supplies feedback, and its curator issues model-mediated changes to retained context (SRC-1, ace/ace.py, ace/core/generator.py, ace/core/curator.py).

prompt = GENERATOR_PROMPT.format(playbook, reflection, question, context) --- ace/core/generator.py

self.playbook, self.next_global_id, operations, _ = self.curator.curate( --- ace/ace.py

The boundary includes both sequential ACE and batched ACEBatch, their model prompts, the optional bulletpoint analyzer, deterministic playbook operations, model-call and logging utilities, and the Finance and two Mind2Web evaluation paths. These callers are part of the target, not excluded host applications: they select the orchestrator, load initial playbook files, and run training or evaluation (SRC-1, eval/finance/run.py, eval/mind2web/run.py, eval/mind2web2/run.py). Persistence is traced from orchestrator output back to caller input and generator consumption (SRC-1, ace/ace.py, utils.py, ace/core/generator.py).

final_playbook_path = os.path.join(save_path, f"final_playbook.txt") with open(final_playbook_path, "w") as f: f.write(self.playbook) --- ace/ace.py:818-820

def load_initial_playbook(path): """Load initial playbook if provided.""" if path and os.path.exists(path): with open(path, 'r') as f: return f.read() return None --- eval/finance/run.py

The batched path is material, not an interchangeable name for the sequential loop. Its inspected update stage generates grouped curator proposals against a shared base playbook, reduces them, then applies the resulting operations (SRC-1, ace/ace_batch.py).

final_operations, _ = self.curator.aggregate_operations( --- ace/ace_batch.py

The boundary kind is complete artifact, partial loop, not whole-system. The frozen main repository is available, but its ace-appworld Gitlink points to a separate repository whose contents are absent from the acquired checkout. The tracked .gitmodules identifies that dependency; the tree pins it to 9f3e92155345a9159f3a8b25abc334eeca05b545 (SRC-1, .gitmodules, ace-appworld). That pointer is not a second frozen source. Its absence prevents implementation conclusions about AppWorld environment interaction, feedback, retained-state consumption and the full AppWorld adaptation loop. The main repository's benchmark claims remain author reports, not verified outcomes (SRC-1, README.md). No additional repository is acquired or silently admitted.

[submodule "ace-appworld"] path = ace-appworld url = https://github.com/ace-agent/ace-appworld.git --- .gitmodules

The applicability cutoff is 2026-10-18. The reviewed boundary is exactly commit 82709de050e1db6e6ef2f07bcb0393560b94992a; the checkout HEAD was checked against acquisition. No claim covers later revisions. The overall evidence tier is code-grounded. Reading implementation is not execution evidence: this run has not executed ACE, tests, examples or model calls. Provider behavior, embedding-model behavior, realized reliability, cost and causal performance gains cannot be established from these inspections. External model services, Python packages and user-supplied data or initial playbooks remain dependencies, not additional evidence sources (SRC-1, pyproject.toml, eval/finance/run.py, ace/core/generator.py).

Repository coverage and remaining inspection

This inventory records the functional scope and initial inspection, not a claim that every included file has already been read. Included but uninspected files remain available at the registered commit for later jobs. They are not access gaps.

Source register