Official implementation of LATTE from "Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs"
Requirements: Python 3.11+
pip install -r requirements.txtCreate a .env file in the repo root with your API key(s):
# Use one or more providers
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
# Select provider and model (defaults to Anthropic / claude-sonnet-4-6)
LLM_PROVIDER=anthropic
LLM_MODEL=claude-sonnet-4-6Supported providers: anthropic, openai.
All commands are run from the repo root.
Run LATTE (dynamic condition) on a single task:
python scripts/run_dynamic_pipeline.py \
--tasks tasks/simple_eda/task_description.json \
--n 4Key options:
| Flag | Default | Description |
|---|---|---|
--tasks |
(required) | Path to task_description.json |
--n |
3 |
Number of teammate agents |
--max-rounds |
40 |
Round limit |
--provider |
env / anthropic |
LLM provider |
--model |
env / provider default | Model name |
--straggler-rounds |
disabled | Automatically release tasks held for N rounds without progress. Otherwise, the lead will receive a heartbeat notification. |
Results are written to runs/<provider>_<model>/<task>_n<N>_<timestamp>/.
| Task | Directory | Description |
|---|---|---|
| Simple EDA | tasks/simple_eda/ |
Analyze an opaque HR dataset; produce config.py, findings.json, summary.txt |
| Debug Sprint | tasks/debug_sprint/ |
Find and fix bugs in a signal-processing library (signal_toolkit.py) |
| Text Processing | tasks/textproc/ |
Extend an existing text processing library with new modules |
All four baseline conditions used in the paper are included:
| Condition | Description |
|---|---|
dynamic |
LATTE — adaptive task graph with straggler detection and verification |
static |
Task graph planned once upfront; no updates during execution |
no_graph |
No task graph; agents coordinate freely via broadcasts and file edits |
decentralized |
No lead agent; all agents self-coordinate without a graph |
metagpt |
MetaGPT multi-agent baseline |
To run a specific baseline directly:
# No-graph baseline
python scripts/run_no_graph_pipeline.py \
--tasks tasks/debug_sprint/task_description.json \
--n 4
# MetaGPT baseline
python scripts/run_metagpt_pipeline.py \
--tasks tasks/textproc/task_description.jsonEach dispatch script runs all conditions for one task and writes results to experiments/:
# Simple EDA (experiment 1)
python scripts/experiment1_dispatch.py --k 10 --n 4 --provider anthropic
# Debug Sprint (experiment 2)
python scripts/experiment2_dispatch.py --k 10 --n 4 --provider anthropic
# Text Processing (experiment 3)
python scripts/experiment_textproc_dispatch.py --k 10 --n 4 --provider anthropicKey options for dispatch scripts:
| Flag | Default | Description |
|---|---|---|
--k |
3 |
Trials per condition |
--n |
4 |
Number of teammate agents |
--conditions |
all | Comma-separated subset, e.g. dynamic,static |
--provider / --model |
env defaults | LLM provider and model |
Trials are interleaved across conditions to reduce time-of-day confounds. Results append to experiments/<task>/n<N>/results.csv so interrupted runs can be resumed.
Open notebooks/run_visualizer_overview.ipynb in Jupyter. The notebook loads all three results.csv files and produces the figures from the paper: pass rates, token and wall-clock costs, coordination metrics, LATTE operator usage, task-graph evolution.
Pre-computed experiment data (results CSVs and aggregated metric JSONs) is included in experiments/.
- Create
tasks/<your_task>/task_description.jsonwithproject,description,success_criteria, andtasks: []fields. - Add
tasks/<your_task>/test_suite.pywith a pytest suite that defines success. - Optionally add
tasks/<your_task>/setup_data.pywith asetup(repo_dir)function to stage data files before agents start. - Run with
python scripts/run_dynamic_pipeline.py --tasks tasks/<your_task>/task_description.json.
MIT
