Multi Agent Kernel (MAK)

---

A kernel for concurrent multi-agent software development.

Multiple agents edit one shared working directory at the same time.

No worktrees, no merge step, no late-stage reconciliation.

The Multi Agent Kernel arbitrates concurrent access the way an OS arbitrates shared memory between threads.

The idea

Most multi-agent coding systems give each agent a Git branch and merge at the end. A message-passing model where conflicts surface late, after the dependency information needed to resolve them is gone.

The Multi Agent Kernel takes the shared-memory approach.

• The codebase is decomposed into independently lockable AST nodes (functions, methods, classes, headers).

• Files on disk are derived artifacts reconstructed from a versioned node store.

• The kernel owns a symbol-level lock table and resolves conflicts at scheduling time, where the dependency graph is still explicit.

• Each agent receives only the nodes it holds write locks on, edits them in isolation, and returns the modified fragments. The kernel reassembles the file.

Check out the knowledge graph for this project. (created with graphify)

Run

⚠️ Currently, MAK only supports Python codebases, there are plans to add other language support in the near future.

1. Clone from source (Python ≥ 3.11)

git clone https://github.com/chaseungjoon/multi-agent-kernel
cd multi-agent-kernel
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

2. Set the API keys

Currently, MAK drives hosted models from three providers — Anthropic, OpenAI, and Google Gemini.

Put the key(s) for the providers you'll use in mak/.env. You only need keys for the agents you actually run:

cp mak/.env.example mak/.env     # Fill in your keys 

3. Choose the number and types of agents & Run

⚠️ Just to be safe, create a separate branch for MAK to work on

# Example with claude 4.8, gpt 5.5 and gemini 3 pro
python3 -m mak --task "your task" --work-dir /path/to/project \
  --models anthropic:claude-opus-4-8 openai:gpt-5.5 gemini:gemini-3-pro

# Example with claude 4.6 X 5
python3 -m mak --task "your task" --work-dir /path/to/project \
  --models anthropic --max-agents 5

Command line arguments

# Describe task
--task "Describe your task here"

# Set working directory
--work-dir /path/to/project

# Omit human review (Not recommended)
--no-review

# Default model
--models anthropic
--models openai
--models gemini

# Set model
--models anthropic:claude-opus-4-8
--models openai:gpt-5.4
--models gemini:gemini-3-pro

# Use multiple providers
--models anthropic openai gemini
--models anthropic:claude-opus-4-8 openai:gpt-5.5 gemini:gemini-3-pro

# Use single provider with multiple agents
--models anthropic --max-agents 5 
--models anthropic:claude-opus-4-8 --max-agents 3

# Choose a custom config file
--config /path/to/config.yaml

Default models list for each provider

For a ready-made target, try the bundled demo:

# Demo MAK
python3 -m mak --task "Implement every function in the dataforge package per its docstring." \
  --config demo/config.yaml --no-review

# Verify the result
python3 -m pytest demo/project

Benchmark

benchmark/ pits MAK against a traditional git-worktree multi-agent workflow on the same workload with the same agents (3× claude-sonnet-4-6). Every operation must edit one shared registry function. The numbers below are the mean of 10 independent runs

• benchmark/project_template_2/ — 90 operations, 9 modules

  	MAK	Git worktrees
  Avg. Tokens	18,339	23,760
  Avg. Time	226.5s	99.5s
  Avg. Accuracy	94% (253.1/270)	93% (251.6/270)
  Avg. Merge conflicts	0	2

MAK spends 23% fewer tokens and hits zero merge conflicts by construction. It also has a slight edge in accuracy.

More statistics

Both sides got a few of the harder algorithms wrong, but the worktree side additionally resulted in 2 merge conflicts.

MAK is slower than traditional worktree based operations because every task contends on that one symbol, so MAK serializes those writes while the worktrees edit in parallel and reconcile afterward: the trade is correctness by construction and token efficiency for execution time on a deliberately maximally-contended workload.

Run it yourself (all targets) with

python3 benchmark/run_benchmark.py --mode real \
  --models anthropic --max-agents 3

Contribute

CONTRIBUTING.md is the full guide — architecture, every subsystem in depth, setup, the quality gates, coding standards, and where to help.

License

MIT © 2026 Seungjoon Cha