SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.07429 · 2024

Agent Workflow Memory

arXiv · PDF · Open in the Atlas

Code that ran

10 of 14 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 8 failures of our instrument; 4 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2409.07429.json · badge: /api/ran/2409.07429.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
zorazrw/agent-workflow-memory canonical (classifier label)
classifier basis: census
10 ran · 4 no recorded run
FunctionStatusWhere it lives
clean_workflow_name Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/mind2web/workflow/retrieve.py
code served (permissive licence) · get_code("1fd693309316e206")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

ARGS = {
    'name': 'workflow: `My Workflow` task'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_think_and_action Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/webarena/induce_prompt.py
code served (permissive licence) · get_code("12b837f02259b5dc")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os

# Create a temporary log file with the expected format
temp_dir = tempfile.mkdtemp()
log_path = os.path.join(temp_dir, "test_log.txt")

# Write a sample log file with the expected structure
# Format: blocks separated by blank lines, odd blocks are "think" blocks, even blocks are "action" blocks
log_content = """First think block
Some context here
browsergym.experiments.loop - INFO - Task step 1: analyze the page

action1()
click('element_id', 42)

Second think block
After first action
browsergym.experiments.loop - INFO - Now we need to fill the form

fill('form_field', 'value')
action2()

Third think block
Continuing analysis
browsergym.experiments.loop - INFO - Final step to complete

click('submit', 1)

"""

with open(log_path, 'w') as f:
    f.write(log_content)

ARGS = {
    "path": log_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_average Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
zorazrw/agent-workflow-memory/mind2web/results/calc_score.py
code served (permissive licence) · get_code("b3e6a3e94633a331")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Construct arguments for get_average function
score_list = [0.5, 0.75, 0.9, 0.85]
percentage = False

ARGS = {
    "score_list": score_list,
    "percentage": percentage,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_data_dict Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/mind2web/offline_induction.py
code served (permissive licence) · get_code("843f258f80209cff")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create temporary JSON files with sample data
temp_dir = tempfile.mkdtemp()

# Create first data file
data1 = [
    {
        'domain': 'tech',
        'subdomain': 'ai',
        'website': 'openai.com',
        'content': 'sample content 1'
    },
    {
        'domain': 'tech',
        'subdomain': 'ai',
        'website': 'deepmind.com',
        'content': 'sample content 2'
    },
    {
        'domain': 'tech',
        'subdomain': 'cloud',
        'website': 'aws.com',
        'content': 'sample content 3'
    }
]

# Create second data file
data2 = [
    {
        'domain': 'social',
        'subdomain': 'messaging',
        'website': 'twitter.com',
        'content': 'sample content 4'
    },
    {
        'domain': 'tech',
        'subdomain': 'ai',
        'website': 'openai.com',
        'content': 'sample content 5'
    }
]

# Write files to temp directory
path1 = os.path.join(temp_dir, 'data1.json')
path2 = os.path.join(temp_dir, 'data2.json')

with open(path1, 'w') as f:
    json.dump(data1, f)

with open(path2, 'w') as f:
    json.dump(data2, f)

# Define ARGS for the function
ARGS = {
    'paths': [path1, path2]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_examples Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/mind2web/workflow/retrieve.py
code served (permissive licence) · get_code("a1075bc76f348b17")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import json
import tempfile
import shutil

# Create a temporary directory with test JSON files
temp_dir = tempfile.mkdtemp()

# Create test data files
test_data_1 = [
    {'website': 'example.com', 'content': 'test1'},
    {'website': 'other.com', 'content': 'test2'},
    {'website': 'example.com', 'content': 'test3'},
]

test_data_2 = [
    {'website': 'example.com', 'content': 'test4'},
    {'website': 'another.com', 'content': 'test5'},
]

# Write test data to files
with open(os.path.join(temp_dir, 'data1.json'), 'w') as f:
    json.dump(test_data_1, f)

with open(os.path.join(temp_dir, 'data2.json'), 'w') as f:
    json.dump(test_data_2, f)

ARGS = {
    'data_dir': temp_dir,
    'website': 'example.com',
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_exemplars Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/mind2web/memory.py
code served (permissive licence) · get_code("0fe1dcf6a2957447")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import json
import tempfile
import shutil

# Create a temporary directory structure for the function to use
tmpdir = tempfile.mkdtemp()
memory_path = os.path.join(tmpdir, 'memory')
os.makedirs(memory_path, exist_ok=True)

# Create a workflow file
workflow_path = os.path.join(tmpdir, 'workflow.txt')
with open(workflow_path, 'w') as f:
    f.write("This is a sample workflow description for testing purposes.")

# Create exemplars.json with sample data
exemplars_data = [
    [
        {
            'role': 'user',
            'content': 'Example 1',
            'specifier': 'domain1_subdomain1_website1'
        },
        {
            'role': 'assistant',
            'content': 'Response 1'
        }
    ],
    [
        {
            'role': 'user',
            'content': 'Example 2',
            'specifier': 'domain1_subdomain1_website2'
        },
        {
            'role': 'assistant',
            'content': 'Response 2'
        }
    ],
    [
        {
            'role': 'user',
            'content': 'Example 3',
            'specifier': 'domain1_subdomain2_website3'
        },
        {
            'role': 'assistant',
            'content': 'Response 3'
        }
    ],
    [
        {
            'role': 'user',
            'content': 'Example 4',
            'specifier': 'domain2_subdomain3_website4'
        },
        {
            'role': 'assistant',
            'content': 'Response 4'
        }
    ]
]

exemplars_path = os.path.join(memory_path, 'exemplars.json')
with open(exemplars_path, 'w') as f:
    json.dump(exemplars_data, f)

# Create a mock args object
class Args:
    def __init__(self):
        self.workflow_path = workflow_path
        self.memory_path = memory_path
        self.domain = 'domain1'
        self.subdomain = 'subdomain1'
        self.website = 'website1'
        self.retrieve_top_k = 2

ARGS = {'args': Args()}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list — a list output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_trajectory Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/mind2web/online_induction.py
code served (permissive licence) · get_code("dfc08c39cb335e67")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary JSON file with the expected structure
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "trajectory.json")

# Create sample data matching the function's expectations
data = [
    {
        "input": [{"content": "print('hello')"}],
        "output": "executed successfully"
    },
    {
        "input": [{"content": "x = 5"}],
        "output": "variable assigned"
    },
    {
        "invalid": "item"  # This will be skipped by is_io_dict
    },
    {
        "input": [{"content": "result = x + 3"}],
        "output": "calculation done"
    }
]

with open(temp_file, 'w') as f:
    json.dump(data, f)

ARGS = {
    "path": temp_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
is_io_dict Ran
ran, violated the contract we drafted
zorazrw/agent-workflow-memory/mind2web/online_induction.py
code served (permissive licence) · get_code("e887f421a98b95cd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Test case 1: valid io_dict with 'input' and 'output' keys
valid_io_dict = {
    'input': torch.randn(2, 4),
    'output': torch.randn(2, 4)
}

# Test case 2: string (not a dict)
string_item = "not a dict"

# We'll test the function with the valid io_dict
ARGS = {
    'item': valid_io_dict
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_workflows Ran
ran, honoured the contract we drafted
zorazrw/agent-workflow-memory/mind2web/workflow/retrieve.py
code served (permissive licence) · get_code("1d36e9593f5fb131")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os

# Create a temporary file with workflow content
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "example_website.md")

workflow_content = """# Workflow: Data Processing
This is a workflow for processing data efficiently.

Some implementation details here.

# Another Workflow: Model Training
Training workflows are important.

More details about training.

# Short Block
Too few.

# Third Workflow: Evaluation
Evaluate model performance here.

Additional evaluation steps.
"""

with open(temp_file, 'w') as f:
    f.write(workflow_content)

ARGS = {
    'path': temp_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
remove_invalid_steps Ran
ran; our contract draft was wrong, not the code
zorazrw/agent-workflow-memory/webarena/induce_prompt.py
code served (permissive licence) · get_code("68200314270ba3be")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

def remove_invalid_steps(actions: list[str]) -> list[str]:
    """Remove invalid steps from the action sequence."""
    valid_actions = []
    for a in actions:
        if 'click(' in a:
            arg = a[a.index('(') + 1:a.index(')')]
            try:
                if type(eval(arg)) == str and type(eval(arg[1:-1])) == int:
                    valid_actions.append(a)
            except:
                continue
        elif 'fill(' in a:
            arg = a[a.index('(') + 1:a.index(',')].strip()
            if type(eval(arg)) == str:
                valid_actions.append(a)
        elif 'scroll(' in a or 'noop(' in a:
            continue
        else:
            valid_actions.append(a)
    return valid_actions

# Test data with various action types
actions = [
    'click("button", 5)',           # valid click: string and int
    'click("link", "text")',        # invalid click: two strings
    'click(123, 456)',              # invalid click: two ints
    'fill("form", value="test")',   # valid fill: has string argument
    'fill(123, value="test")',      # invalid fill: int argument
    'scroll(5)',                    # scroll should be filtered
    'noop()',                       # noop should be filtered
    'custom_action()',              # valid else: passes through
]

ARGS = {
    "actions": actions
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_think_and_action Not yet run
unverified (harvested, no recorded run)
zorazrw/agent-workflow-memory/webarena/induce_rule.py
code served (permissive licence) · get_code("9d59295679823ebc")
How it ran

no invocation recorded.

get_split Not yet run
unverified (harvested, no recorded run)
zorazrw/agent-workflow-memory/mind2web/offline_induction.py
code served (permissive licence) · get_code("1c21d96a59bbdec1")
How it ran

no invocation recorded.

load_blocks Not yet run
unverified (harvested, no recorded run)
zorazrw/agent-workflow-memory/webarena/induce_prompt.py
code served (permissive licence) · get_code("868712f71f85de88")
How it ran

no invocation recorded.

remove_invalid_steps Not yet run
unverified (harvested, no recorded run)
zorazrw/agent-workflow-memory/webarena/induce_rule.py
code served (permissive licence) · get_code("5f0fd52ba91452d0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 2 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 2 later) and 11 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
load_blocks 1 other paper
1 later
2509.25140 (2025-09)
Re as on i ng Ban k: Scaling Agent Self-Evolving with Reasoning Memory
remove_invalid_steps 1 other paper
1 later
2509.25140 (2025-09)
Re as on i ng Ban k: Scaling Agent Self-Evolving with Reasoning Memory

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. To build agents that can similarly benefit from this process, we introduce Agent Workflow Memory (AWM), a method for inducing commonly reused routines, i.e., workflows, and selectively providing workflows to the agent to guide subsequent generations. AWM flexibly applies to both offline and online scenarios, where agents induce workflows from training examples beforehand or from test queries on the fly. We experiment on two major web navigation benchmarks -- Mind2Web and WebArena -- that collectively cover 1000+ tasks from 200+ domains across travel, shopping, and social media, among others. AWM substantially improves the baseline results by 24.6% and 51.1% relative success rate on Mind2Web and WebArena while reducing the number of steps taken to solve WebArena tasks successfully. Furthermore, online AWM robustly generalizes in cross-task, website, and domain evaluations, surpassing baselines from 8.9 to 14.0 absolute points as train-test task distribution gaps widen.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2409.07429")
get_code_for_paper("2409.07429")
have("2409.07429")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2409.07429.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2409.07429.svg)](https://syntology.ai/paper/2409.07429)

Connect an agent — have() is free.