MegaDeepagents
GitHub
OPEN-SOURCE AGENT RUNTIME / V3

One runtime.
Accountable agent teams.

MegaDeepagents turns a goal into governed, recoverable work—from one agent to a coordinated team—through a single LangGraph control plane.

Explore on GitHub See how it works
MIT licensed Model agnostic Self-hosted
runtime.map V3 / READY
00 GOAL Ship a verified change
01 SUPERVISOR route + govern
02A SINGLE focused path
02B TEAM parallel tasks
03 VERIFY evidence gate
14:08:31 plan.created 6 tasks
14:08:34 agents.dispatched 3 workers
14:09:02 artifact.verified PASS
01production Root Graph
35typed API paths
456passing backend tests
100%open source
01 / WHY

Agents should not mark their own homework.

Demos stop when the model returns text. Production work begins there. MegaDeepagents separates planning, execution and verification so every successful run has durable state, evidence and a recoverable path.

01

Workers produce artifacts. Verifiers decide success.

02

Single and team runs share one execution model.

03

Every decision can be replayed, inspected and resumed.

02 / CAPABILITIES

A control plane for work that has to finish.

The runtime coordinates autonomy without giving up control, evidence or recovery.

01

One graph, every run

Complexity changes the route—not the runtime. Focused tasks and parallel teams use the same state, checkpoints and completion rules.

02
↻

Durable by default

SQLite WAL, event envelopes and LangGraph checkpoints make pause, resume and process recovery first-class behavior.

03
✓

Evidence before success

Artifacts, tests and output contracts pass through a fail-closed verification gate before a task can complete.

04
⌁

Isolated code work

Governed Git worktrees give each worker an isolated workspace before controlled integration.

05
◆

Human approval where it matters

Plans, permissions and risky operations can pause at explicit human-in-the-loop checkpoints.

06
≋

Observable from the source

Typed events, SSE replay, task graphs, agent messages and artifact lineage expose what happened and why.

03 / ARCHITECTURE

Simple boundaries. Serious execution.

LangGraph owns orchestration. TaskGraph owns the plan. TaskBoard owns execution state. Workers create evidence. Verifier alone owns success.

Read the architecture guide
ENTRY API V1 · CLI · Runtime Console
↓
CONTROL LangGraph Root Graph Supervisor · HITL · Replan · Finalize
↓
PLAN TaskGraph dependencies + contracts
STATE TaskBoard claims + attempts + gates
↓
WORK DeepAgents tools + worktrees
PROOF Verifier artifacts + evidence
↓
DURABLE SQLite · Checkpoints · Event Replay
04 / EXECUTION FLOW

From intent to verified outcome.

  1. 01

    Understand

    Normalize the goal, constraints, repository and completion contract.

  2. 02

    Route

    Choose a focused agent or construct a role-aware team.

  3. 03

    Execute

    Claim dependency-ready tasks and work in isolated environments.

  4. 04

    Verify

    Test outputs, inspect evidence and create real repair work when needed.

  5. 05

    Recover

    Pause, approve, resume or replay without losing the run.

05 / QUICKSTART

Run it where your work lives.

MegaDeepagents is self-hosted by design. Start locally, bring your own model, and keep repositories and runtime state under your control.

Python 3.11+ Node.js 20+ Git
quickstart.sh 
# clone & configure
git clone https://github.com/xxiaoxiong/MegaDeepagents.git
cd MegaDeepagents
cp .env.example .env

# launch the durable runtime
docker compose up --build
API, runtime console and durable state—ready.
OPEN SOURCE / MIT

Build agents you can trust to finish.

View the repository Read the docs