OpenDevin Architecture

This directory contains the core components of OpenDevin.

Classes

The key classes in OpenDevin are:

• LLM: brokers all interactions with large language models. Works with any underlying completion model, thanks to LiteLLM.
• Agent: responsible for looking at the current State, and producing an Action that moves one step closer toward the end-goal.
• AgentController: initializes the Agent, manages State, and drive the main loop that pushes the Agent forward, step by step
• State: represents the current state of the Agent's task. Includes things like the current step, a history of recent events, the Agent's long-term plan, etc
• EventStream: a central hub for Events, where any component can publish Events, or listen for Events published by other components
  • Event: an Action or Observeration
    • Action: represents a request to e.g. edit a file, run a command, or send a message
    • Observation: represents information collected from the environment, e.g. file contents or command output
• Runtime: responsible for performing Actions, and sending back Observations
  • Sandbox: the part of the runtime responsible for running commands, e.g. inside of Docker
• Server: brokers OpenDevin sessions over HTTP, e.g. to drive the frontend
  • Session: holds a single EventStream, a single AgentController, and a single Runtime. Generally represents a single task (but potentially including several user prompts)
  • SessionManager: keeps a list of active sessions, and ensures requests are routed to the correct Session

Control Flow

Here's the basic loop (in pseudocode) that drives agents.

while True:
  prompt = agent.generate_prompt(state)
  response = llm.completion(prompt)
  action = agent.parse_response(response)
  observation = runtime.run(action)
  state = state.update(action, observation)

In reality, most of this is achieved through message passing, via the EventStream. The EventStream serves as the backbone for all communication in OpenDevin.

flowchart LR
  Agent--Actions-->AgentController
  AgentController--State-->Agent
  AgentController--Actions-->EventStream
  EventStream--Observations-->AgentController
  Runtime--Observations-->EventStream
  EventStream--Actions-->Runtime
  Frontend--Actions-->EventStream

Loading

Runtime

The Runtime class is abstract, and has a few different implementations:

• We have a LocalRuntime, which runs commands and edits files directly on the user's machine
• We have a DockerRuntime, which runs commands inside of a docker sandbox, and edits files directly on the user's machine
• We have an E2BRuntime, which uses e2b.dev containers to sandbox file and command operations