Parallel Benchmarking with Git Worktrees

Overview

Modern agentic workflows often involve multiple specialized agents working on different parts of a codebase simultaneously. When these agents modify files, they risk stepping on each other’s toes—causing merge conflicts or corrupted git states.

HydraR addresses this by leveraging Git Worktrees for isolated execution. This guide demonstrates a parallel workflow where: 1. Three Agents simultaneously write R code for different sorting algorithms (Bubble, Quick, Merge sort) in separate worktrees. 2. A Merge Harmonizer automatically merges these independent branches back into the main repository. 3. A Benchmarking Node executes the generated code to compare their performance. 4. A Visualization Node plots the results.

---

🏗️ Step 1: Initialize an Isolated Repository

To demonstrate the power of worktrees, we first create a temporary Git repository.

library(HydraR)
library(withr)
library(ggplot2)
library(future)
library(furrr)

# Check for Anthropic API Key (Needed for refined agents)
# Must be done BEFORE future::plan so workers inherit the environment
if (Sys.getenv("ANTHROPIC_API_KEY") == "") {
  if (file.exists("../.Renviron")) readRenviron("../.Renviron")
  if (file.exists(".Renviron")) readRenviron(".Renviron")
}

if (Sys.getenv("ANTHROPIC_API_KEY") == "") {
  stop("ANTHROPIC_API_KEY not found. Please set it in your .Renviron or workspace.")
}

# Ensure Gemini CLI path is configured (Environment variables are inherited by workers)
Sys.setenv(HYDRAR_GEMINI_PATH = "/opt/homebrew/bin/gemini")

# 0. Setup Parallel Execution (for Worktree Isolation)
future::plan(future::multisession, workers = 3)
options(future.rng.onMisuse = "ignore")

# 1. Create a temporary folder
repo_root <- file.path(tempdir(), "sorting-benchmark-repo")
if (dir.exists(repo_root)) unlink(repo_root, recursive = TRUE)
dir.create(repo_root)

# 2. Initialize Git and a README
withr::with_dir(repo_root, {
  system2("git", c("init"))
  system2("git", c("config", "user.email", "apaf@example.com"))
  system2("git", c("config", "user.name", "APAF Agent"))
  system2("git", c("config", "commit.gpgsign", "false"))
  writeLines("# Sorting Benchmark", "README.md")
  system2("git", c("add", "README.md"))
  system2("git", c("commit", "-m", "Initial_commit"))
  system2("git", c("branch", "-M", "main"))
})

---

🧠 Step 2: Declarative Workflow Loading

Rather than registering roles and logic manually in R, we now define the entire workflow architecture (graph, prompts, logic) in a single declarative YAML file. This follows the APAF-Agentic-Standard for Zero-R-Code definitions.

# 1. Load the workflow manifest
wf <- load_workflow("sorting_benchmark.yml")

# 2. Extract components
mermaid_graph <- wf$graph
initial_state <- wf$initial_state

# 3. Inject dynamic environment variables into initial state
initial_state$repo_root <- repo_root
initial_state$output_dir <- file.path(getwd(), "..", "paper", "figures")

# 4. Instantiate and Compile
dag <- AgentDAG$from_mermaid(mermaid_graph, node_factory = auto_node_factory())
compiled_dag <- dag$compile()

# 5. Run with Worktree Isolation
# Ensure workers inherit our environment (for API keys and PATH)
future::plan(future::multisession, workers = 3)

results <- compiled_dag$run(
  initial_state = initial_state,
  use_worktrees = TRUE,
  repo_root = repo_root,
  fail_if_dirty = FALSE,
  packages = c("withr", "HydraR", "ggplot2")
)

# 6. Save Execution Trace
compiled_dag$save_trace("sorting_trace.json")

---

🧘 Summary

By using Declarative Mermaid Nodes with auto_node_factory(), we: 1. Eliminated Boilerplate: No hand-written node factory — the Mermaid graph IS the specification. 2. Eliminated State Corruption: Three agents modified the codebase simultaneously via isolated Git Worktrees. 3. Automated Conflict Resolution: The MergeHarmonizer handled integration of branches. 4. End-to-End Validation: Generated code was benchmarked immediately downstream.

---