24b71927c3
* Some improvements to prompts, some better exception handling for various file IO errors, added timeout and max return token configurations for the LLM api. * More monologue prompt improvements * Dynamically set username provided in prompt. * Remove absolute paths from llm prompts, fetch working directory from sandbox when resolving paths in fileio operations, add customizable timeout for bash commands, mention said timeout in llm prompt. * Switched ssh_box to disabling tty echo and removed the logic attempting to delete it from the response afterwards, fixed get_working_directory for ssh_box. * Update prompts in integration tests to match monologue agent changes. * Minor tweaks to make merge easier. * Another minor prompt tweak, better invalid json handling. * Fix lint error * More catch-up to fix lint errors introduced by merge. * Force WORKSPACE_MOUNT_PATH_IN_SANDBOX to match WORKSPACE_MOUNT_PATH in local sandbox mode, combine exception handlers in prompts.py. --------- Co-authored-by: Jim Su <jimsu@protonmail.com> Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>
Agent Framework Research
In this folder, there may exist multiple implementations of Agent that will be used by the framework.
For example, agenthub/monologue_agent, agenthub/metagpt_agent, agenthub/codeact_agent, etc.
Contributors from different backgrounds and interests can choose to contribute to any (or all!) of these directions.
Constructing an Agent
The abstraction for an agent can be found here.
Agents are run inside of a loop. At each iteration, agent.step() is called with a
State input, and the agent must output an Action.
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.
State
The state contains:
- A history of actions taken by the agent, as well as any observations (e.g. file content, command output) from those actions
- A list of actions/observations that have happened since the most recent step
- A
plan, which contains the main goal- The agent can add and modify subtasks through the
AddTaskActionandModifyTaskAction
- The agent can add and modify subtasks through the
Actions
Here is a list of available Actions, which can be returned by agent.step():
CmdRunAction- Runs a command inside a sandboxed terminalCmdKillAction- Kills a background commandFileReadAction- Reads the content of a fileFileWriteAction- Writes new content to a fileBrowseURLAction- Gets the content of a URLAgentRecallAction- Searches memory (e.g. a vector database)AddTaskAction- Adds a subtask to the planModifyTaskAction- Changes the state of a subtaskAgentThinkAction- A no-op that allows the agent to add plaintext to the history (as well as the chat log)AgentFinishAction- 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:
CmdOutputObservationBrowserOutputObservationFileReadObservationFileWriteObservationUserMessageObservationAgentRecallObservationAgentErrorObservation
You can use observation.to_dict() and observation_from_dict to serialize and deserialize observations.
Interface
Every agent must implement the following methods:
step
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.
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.