Update agent docs (#407)

* Update agent docs

* Update README.md

* Update README.md
This commit is contained in:
Robert Brennan
2024-03-30 19:23:40 -04:00
committed by GitHub
parent 2d8cb973a8
commit 93cb5e0617

View File

@@ -9,9 +9,52 @@ Contributors from different backgrounds and interests can choose to contribute t
The abstraction for an agent can be found [here](../opendevin/agent.py).
On a high-level, at each step, an agent takes in a [State](../opendevin/state.py) object and outputs an [Action](../opendevin/action).
Agents are run inside of a loop. At each iteration, `agent.step()` is called with a
[State](../opendevin/state.py) input, and the agent must output an [Action](../opendevin/action).
Your agent must implement the following methods:
Every agent also has a `self.llm` which it can use to interact with the LLM configured by the user.
See the [LiteLLM docs for `self.llm.completion`](https://docs.litellm.ai/docs/completion).
## State
The `state` contains:
* A history of actions taken by the agent, as well as any obserations (e.g. file content, command output) from those actions
* A list of actions/observations that have happened since the most recent step
* A [`plan`](https://github.com/OpenDevin/OpenDevin/blob/main/opendevin/plan.py), which contains the main goal
* The agent can add and modify subtasks through the `AddTaskAction` and `ModifyTaskAction`
## Actions
Here is a list of available Actions, which can be returned by `agent.step()`:
- [`CmdRunAction`](../opendevin/action/bash.py) - Runs a command inside a sandboxed terminal
- [`CmdKillAction`](../opendevin/action/bash.py) - Kills a background command
- [`FileReadAction`](../opendevin/action/fileop.py) - Reads the content of a file
- [`FileWriteAction`](../opendevin/action/fileop.py) - Writes new content to a file
- [`BrowseURLAction`](../opendevin/action/browse.py) - Gets the content of a URL
- [`AgentRecallAction`](../opendevin/action/agent.py) - Searches memory (e.g. a vector database)
- [`AddTaskAction`](../opendevin/action/tasks.py) - Adds a subtask to the plan
- [`ModifyTaskAction`](../opendevin/action/tasks.py) - Changes the state of a subtask
- [`AgentThinkAction`](../opendevin/action/agent.py) - A no-op that allows the agent to add plaintext to the history (as well as the chat log)
- [`AgentFinishAction`](../opendevin/action/agent.py) - Stops the control loop, allowing the user to enter a new task
You can use `action.to_dict()` and `action_from_dict` to serialize and deserialize actions.
## Observations
There are also several types of Observations. These are typically available in the step following the corresponding Action.
But they may also appear as a result of asynchronous events (e.g. a message from the user, logs from a command running
in the background).
Here is a list of available Observations:
- [`CmdOutputObservation`](../opendevin/observation/run.py)
- [`BrowserOutputObservation`](../opendevin/observation/browse.py)
- [`FileReadObservation`](../opendevin/observation/files.py)
- [`FileWriteObservation`](../opendevin/observation/files.py)
- [`UserMessageObservation`](../opendevin/observation/)
- [`AgentRecallObservation`](../opendevin/observation/recall.py)
- [`AgentErrorObservation`](../opendevin/observation/error.py)
You can use `observation.to_dict()` and `observation_from_dict` to serialize and deserialize observations.
## Interface
Every agent must implement the following methods:
### `step`
```
@@ -20,21 +63,11 @@ def step(self, state: "State") -> "Action"
`step` moves the agent forward one step towards its goal. This probably means
sending a prompt to the LLM, then parsing the response into an `Action`.
We now have [two main categories of actions](../opendevin/action/base.py):
- `ExecutableAction`: will produces a corresponding `Observation` (source [here](../opendevin/observation.py)) for the agent to take the next `Action`.
- `NotExecutableAction`: will produces a `NullObservation` by the [controller](../opendevin/controller/__init__.py), which could means telling the agent to ignore this action.
For `ExecutableAction`, we currently have:
- `CmdRunAction` and `CmdKillAction` for bash command (see source [here](../opendevin/action/bash.py)).
- `FileReadAction` and `FileWriteAction` for file operations (see source [here](../opendevin/action/fileop.py)).
- `BrowseURLAction` to open a web page (see source [here](../opendevin/action/browse.py)).
- `AgentThinkAction`, `AgentFinishAction`: these are non-executable actions for agent to update its status to the user. For example, agent could use `AgentThink` to explain its though process to the user (see source [here](../opendevin/action/agent.py)).
- `AgentEchoAction`: the agent can produce some messages as its own Observation in the next `.step`, this will produces a `AgentMessageObservation` (see source [here](../opendevin/action/agent.py)).
- `AgentRecallAction`: recalls a past memory (see source [here](../opendevin/action/agent.py)).
### `search_memory`
```
def search_memory(self, query: str) -> List[str]:
```
`search_memory` should return a list of events that match the query. This will be used
for the `recall` action.
You can optionally just return `[]` for this method, meaning the agent has no long-term memory.