JAZ boundary

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

Boundary and evidence

This run analyses how JAZ turns keyword inputs and Python tools into model-driven REPL execution, controls recursive invocations, returns results, and records or reuses execution state. JAZ qualifies as an agent system and is classified as an enclosing runtime: its own agent loop queries the model, parses the response, executes committed code, feeds observations back, and decides whether to continue or finish. It is not classified as merely a runtime client because an application can call invoke. SRC-1, src/jaz/invoke.py, src/jaz/_agent.py, and src/jaz/repl/python_repl.py establish these responsibilities.

The source's declared loop is:

Agents execute in a REPL loop where the LLM generates code, observes results, and iterates until the task is complete. --- README.md

The inspected implementation connects model queries to execution through do_one_repl_iteration, including this committed-code call:

exec_result = self.repl_template.exec( state, span.committed_code, str(iteration), exec_timeout_override ) --- src/jaz/_agent.py

The whole-system boundary is the package shipped in this repository, including its console driver, prompts, configuration and credentials, hook governance, recording/replay, and support utilities. It is not a claim to cover the authors' unshipped development or evaluation environment. Both installed jaz and python -m jaz lead to jaz.console.main; the console loads settings, builds its namespace, activates selected hooks, and dispatches interactive, command, file or piped input. The console is included, not treated as an external host. SRC-1, pyproject.toml, src/jaz/__main__.py, src/jaz/console.py.

[project.scripts] jaz = "jaz.console:main" --- pyproject.toml

State-producing and consuming paths are included together. Public invocation builds an Agent and recursive invoke tool; the loop owns per-invocation REPL history and message observations. The console records completed or failed prompt turns and supplies prior live input/result objects to later invocations. Its model-setting writer and startup loader are also inside the boundary. Trajectory recording, trace loading and replay, workflow-package materialization, and return validation are included rather than dismissed as auxiliary callers. SRC-1, src/jaz/invoke.py, src/jaz/_agent.py, src/jaz/console.py, src/jaz/hooks/builtin/atif_trace.py, src/jaz/hooks/builtin/replay.py, src/jaz/hooks/builtin/workflow_replay.py, src/jaz/hooks/builtin/return_hooks.py.

if history.turns: extra["conversation_history"] = history try: result = jaz.invoke(hooks, task=task, *extra) --- src/jaz/console.py

Tracked top-level coverage

The frozen Git tree contains the following five top-level entries. “Included” fixes functional scope; it does not mean every implementation line has already been inspected. Initial reads and bounded searches establish the entry points and material state loop. Files not yet individually read remain available to later jobs under SRC-1, not access gaps.

The repeated src/ row expands one tracked directory; it is not an additional source. No tracked test, benchmark, dataset or evaluator directory is present in this frozen tree. The README explains the publication boundary:

This repository is generated: each release publishes a cleaned copy of the package, so changes are not made here. Contributor setup and code conventions live with the source. --- README.md

This limits conclusions about development-time optimization and evaluation. Shipped validators and replay hooks remain analysable, but absent evaluation tooling and results cannot establish benchmark performance, training gains, or the authors' complete improvement loop. SRC-1, README.md, pyproject.toml, src/jaz/hooks/builtin/return_hooks.py, src/jaz/hooks/builtin/replay.py.

Frozen applicability and external dependencies

The reviewed boundary is exactly commit 0803d4971be785e95b80054b02259664d70fa3da; the checkout's HEAD matched the acquired pin. The applicability cutoff is 2026-10-19, the supplied run date. No later repository changes, live documentation, service responses or unregistered sources are evidence. This is a code-grounded analysis of the published implementation, supplemented by its own documentation. No target code, tests, examples or model calls were executed, and no observed-run or causal-experiment evidence was supplied.

External dependencies include the Python runtime, declared Python packages, provider services, optional telemetry services and user-supplied tools or data. JAZ's adapters and configuration for them are included; their implementations, actual service behavior and arbitrary consuming applications are excluded. Consequently the analysis cannot establish provider reliability, tool-side effects in a real application, sandbox resistance under attack, or end-to-end task success. The default LiteLLM adapter visibly delegates to litellm.completion/litellm.acompletion; that delegation is implementation evidence, not an observed provider call. SRC-1, pyproject.toml, src/jaz/llm/_litellm.py, README.md.

Source register