Reflexion boundary

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

Boundary and evidence

This run analyses Reflexion as shipped experiment workflows for repeated model-dependent task attempts, feedback, self-reflection and subsequent attempts. The intended use is to characterize their control and retained-state mechanisms, not to certify the paper's results or the behavior of a later deployment. The target class is workflow: the repository supplies task-specific drivers and agents rather than a single general-purpose agent service. Its own description establishes the research-artifact context (SRC-1, README.md):

This repo holds the code, demos, and log files for Reflexion: Language Agents with Verbal Reinforcement Learning by Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao. --- README.md

The functional boundary includes all four shipped task families: ALFWorld, HotPotQA, programming and WebShop. It includes their launch scripts or notebooks, agent/model adapters, prompt construction, reflection production, feedback/evaluation code, retained-state writes, reload/resume paths and result consumers. Shipped drivers are part of the target, not excluded host applications. For example, ALFWorld's driver owns the transition from trial execution to reflection updates and persistent environment configuration (SRC-1, alfworld_runs/main.py):

run trial

run_trial(trial_log_path, world_log_path, trial_idx, env_configs, args.use_memory, args.model)

update memory if needed

if args.use_memory: env_configs: List[Dict[str, Any]] = update_memory(trial_log_path, env_configs)

log env configs for trial

with open(trial_env_configs_log_path, 'w') as wf: json.dump(env_configs, wf, indent=4) --- alfworld_runs/main.py

HotPotQA's three notebooks construct the agents, retry questions and save logs and serialized agents; they are included alongside hotpotqa_runs/agents.py and hotpotqa_runs/util.py. Programming's CLI dispatches simple, Reflexion, immediate-reflexion, immediate-refinement, Reflexion-UCS and test-accuracy strategies; the implemented strategy selection, not just the README's examples, defines that branch's scope (SRC-1, programming_runs/main.py). WebShop includes its shipped HTTP environment client, trial runner and memory/logging driver; storing reflections in its history object does not by itself establish that its action-model prompt consumes them (SRC-1, webshop_runs/webshop_trial.py). Later mechanism analysis must trace the actual consumption path separately for each family.

The boundary kind is complete artifact, partial loop. The full parent repository revision is frozen and available, but two material shipped integration paths terminate in empty Git-linked dependency directories. The tree records programming_runs/lazzzy at gitlink commit 404c06a5bf6201596b45d28f092aefe017306cd9 and programming_runs/executors/leetcode_env at 228163abdc983712bebfd8e26f7e7d360830e648. Those are dependency pointers, not additional registered sources. No dependency checkout was acquired or inspected. The declared repositories are visible in SRC-1, .gitmodules:

[submodule "programming_runs/lazzzy"] path = programming_runs/lazzzy url = https://github.com/GammaTauAI/lazzzy [submodule "programming_runs/executors/leetcode_env"] path = programming_runs/executors/leetcode_env url = https://github.com/GammaTauAI/leetcode-hard-gym --- .gitmodules

The UCS strategy imports its search driver from the former dependency (SRC-1, programming_runs/reflexion_ucs.py):

from lazzzy.ucs import ucs --- programming_runs/reflexion_ucs.py

Its callback construction can be analysed, but the missing search implementation prevents a complete account of UCS scheduling and termination. The latter dependency supplies LeetCode environment, submission and formatting APIs referenced by the executor factory and evaluation scripts; its absence prevents a complete account of remote submission, grading and formatting semantics (SRC-1, programming_runs/executors/factory.py, programming_runs/evaluate_leet_results.py, programming_runs/evaluate_rs_leet_results.py). These limits do not block analysis of the parent repository's ordinary reflection loops.

Repository coverage

The table classifies every top-level tracked entry. “Included” establishes functional scope, not a claim that every file has already been inspected. Paths not yet inspected remain available to subsequent source-grounded jobs at the same revision.

All source-dependent coverage above refers to SRC-1 and its commit-relative paths. External execution dependencies include model-provider services and model weights, LangChain/Wikipedia retrieval, the ALFWorld environment and data, the WebShop HTTP server, Python/Rust execution facilities and the two Git-linked programming dependencies. The repository's adapters are included; dependency internals and live service state are excluded or unavailable. Consequently this analysis cannot establish provider output behavior, live retrieval contents, environment/server correctness, deployment availability or secure isolation of generated code merely from these adapters (SRC-1, alfworld_runs/alfworld_trial.py, hotpotqa_runs/agents.py, webshop_runs/webshop_trial.py, programming_runs/generators/model.py, programming_runs/executors/py_executor.py, programming_runs/executors/rs_executor.py).

The frozen revision is 218cf0ef1df84b05ce379dd4a8e47f17766733a0, verified as the acquired checkout's HEAD. The applicability cutoff is 2026-10-14, the supplied run date, not a claim about the commit's publication date. Overall evidence is code-grounded: source text establishes implemented paths; documentation supplies declared intent. No target code, examples or tests were executed. Tracked logs and serialized results are within the frozen parent source but have not been examined as run evidence in this boundary. Their presence and the README's description do not establish a reproduced performance result or causal benefit from reflection.

Source register