diff options
Diffstat (limited to 'autogpts/autogpt/autogpt/agents')
| -rw-r--r--[l---------] | autogpts/autogpt/autogpt/agents/README.md | 38 | ||||
| -rw-r--r-- | autogpts/autogpt/autogpt/agents/agent.py | 43 | ||||
| -rw-r--r-- | autogpts/autogpt/autogpt/agents/base.py | 15 | ||||
| -rw-r--r-- | autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py | 42 |
4 files changed, 83 insertions, 55 deletions
diff --git a/autogpts/autogpt/autogpt/agents/README.md b/autogpts/autogpt/autogpt/agents/README.md index adc158547..4ab573243 120000..100644 --- a/autogpts/autogpt/autogpt/agents/README.md +++ b/autogpts/autogpt/autogpt/agents/README.md @@ -1 +1,37 @@ -../../../../docs/content/AutoGPT/component agent/agents.md
\ No newline at end of file +# 🤖 Agents + +Agent is composed of [🧩 Components](./components.md) and responsible for executing pipelines and some additional logic. The base class for all agents is `BaseAgent`, it has the necessary logic to collect components and execute protocols. + +## Important methods + +`BaseAgent` provides two abstract methods needed for any agent to work properly: +1. `propose_action`: This method is responsible for proposing an action based on the current state of the agent, it returns `ThoughtProcessOutput`. +2. `execute`: This method is responsible for executing the proposed action, returns `ActionResult`. + +## AutoGPT Agent + +`Agent` is the main agent provided by AutoGPT. It's a subclass of `BaseAgent`. It has all the [Built-in Components](./built-in-components.md). `Agent` implements the essential abstract methods from `BaseAgent`: `propose_action` and `execute`. + +## Building your own Agent + +The easiest way to build your own agent is to extend the `Agent` class and add additional components. By doing this you can reuse the existing components and the default logic for executing [⚙️ Protocols](./protocols.md). + +```py +class MyComponent(AgentComponent): + pass + +class MyAgent(Agent): + def __init__( + self, + settings: AgentSettings, + llm_provider: ChatModelProvider, + file_storage: FileStorage, + legacy_config: Config, + ): + # Call the parent constructor to bring in the default components + super().__init__(settings, llm_provider, file_storage, legacy_config) + # Add your custom component + self.my_component = MyComponent() +``` + +For more customization, you can override the `propose_action` and `execute` or even subclass `BaseAgent` directly. This way you can have full control over the agent's components and behavior. Have a look at the [implementation of Agent](https://github.com/Significant-Gravitas/AutoGPT/tree/master/autogpts/autogpt/autogpt/agents/agent.py) for more details. diff --git a/autogpts/autogpt/autogpt/agents/agent.py b/autogpts/autogpt/autogpt/agents/agent.py index 3572cbed0..4a66a7ca4 100644 --- a/autogpts/autogpt/autogpt/agents/agent.py +++ b/autogpts/autogpt/autogpt/agents/agent.py @@ -19,7 +19,6 @@ from autogpt.components.event_history import EventHistoryComponent from autogpt.core.configuration import Configurable from autogpt.core.prompting import ChatPrompt from autogpt.core.resource.model_providers import ( - AssistantChatMessage, AssistantFunctionCall, ChatMessage, ChatModelProvider, @@ -27,7 +26,7 @@ from autogpt.core.resource.model_providers import ( ) from autogpt.core.runner.client_lib.logging.helpers import dump_prompt from autogpt.file_storage.base import FileStorage -from autogpt.llm.providers.openai import get_openai_command_specs +from autogpt.llm.providers.openai import function_specs_from_commands from autogpt.logs.log_cycle import ( CURRENT_CONTEXT_FILE_NAME, NEXT_ACTION_FILE_NAME, @@ -46,7 +45,6 @@ from autogpt.utils.exceptions import ( AgentException, AgentTerminated, CommandExecutionError, - InvalidArgumentError, UnknownCommandError, ) @@ -104,7 +102,11 @@ class Agent(BaseAgent, Configurable[AgentSettings]): self.ai_profile = settings.ai_profile self.directives = settings.directives prompt_config = OneShotAgentPromptStrategy.default_configuration.copy(deep=True) - prompt_config.use_functions_api = settings.config.use_functions_api + prompt_config.use_functions_api = ( + settings.config.use_functions_api + # Anthropic currently doesn't support tools + prefilling :( + and self.llm.provider_name != "anthropic" + ) self.prompt_strategy = OneShotAgentPromptStrategy(prompt_config, logger) self.commands: list[Command] = [] @@ -172,7 +174,7 @@ class Agent(BaseAgent, Configurable[AgentSettings]): task=self.state.task, ai_profile=self.state.ai_profile, ai_directives=directives, - commands=get_openai_command_specs(self.commands), + commands=function_specs_from_commands(self.commands), include_os_info=self.legacy_config.execute_local_commands, ) @@ -202,12 +204,9 @@ class Agent(BaseAgent, Configurable[AgentSettings]): ] = await self.llm_provider.create_chat_completion( prompt.messages, model_name=self.llm.name, - completion_parser=self.parse_and_validate_response, - functions=( - get_openai_command_specs(self.commands) - if self.config.use_functions_api - else [] - ), + completion_parser=self.prompt_strategy.parse_response_content, + functions=prompt.functions, + prefill_response=prompt.prefill_response, ) result = response.parsed_result @@ -223,28 +222,6 @@ class Agent(BaseAgent, Configurable[AgentSettings]): return result - def parse_and_validate_response( - self, llm_response: AssistantChatMessage - ) -> OneShotAgentActionProposal: - parsed_response = self.prompt_strategy.parse_response_content(llm_response) - - # Validate command arguments - command_name = parsed_response.use_tool.name - command = self._get_command(command_name) - if arg_errors := command.validate_args(parsed_response.use_tool.arguments)[1]: - fmt_errors = [ - f"{'.'.join(str(p) for p in f.path)}: {f.message}" - if f.path - else f.message - for f in arg_errors - ] - raise InvalidArgumentError( - f"The set of arguments supplied for {command_name} is invalid:\n" - + "\n".join(fmt_errors) - ) - - return parsed_response - async def execute( self, proposal: OneShotAgentActionProposal, diff --git a/autogpts/autogpt/autogpt/agents/base.py b/autogpts/autogpt/autogpt/agents/base.py index cf8e3cac8..515515701 100644 --- a/autogpts/autogpt/autogpt/agents/base.py +++ b/autogpts/autogpt/autogpt/agents/base.py @@ -39,11 +39,12 @@ from autogpt.core.configuration import ( SystemSettings, UserConfigurable, ) -from autogpt.core.resource.model_providers import AssistantFunctionCall -from autogpt.core.resource.model_providers.openai import ( - OPEN_AI_CHAT_MODELS, - OpenAIModelName, +from autogpt.core.resource.model_providers import ( + CHAT_MODELS, + AssistantFunctionCall, + ModelName, ) +from autogpt.core.resource.model_providers.openai import OpenAIModelName from autogpt.models.utils import ModelWithSummary from autogpt.prompts.prompt import DEFAULT_TRIGGERING_PROMPT @@ -56,8 +57,8 @@ P = ParamSpec("P") class BaseAgentConfiguration(SystemConfiguration): allow_fs_access: bool = UserConfigurable(default=False) - fast_llm: OpenAIModelName = UserConfigurable(default=OpenAIModelName.GPT3_16k) - smart_llm: OpenAIModelName = UserConfigurable(default=OpenAIModelName.GPT4) + fast_llm: ModelName = UserConfigurable(default=OpenAIModelName.GPT3_16k) + smart_llm: ModelName = UserConfigurable(default=OpenAIModelName.GPT4) use_functions_api: bool = UserConfigurable(default=False) default_cycle_instruction: str = DEFAULT_TRIGGERING_PROMPT @@ -174,7 +175,7 @@ class BaseAgent(Configurable[BaseAgentSettings], metaclass=AgentMeta): llm_name = ( self.config.smart_llm if self.config.big_brain else self.config.fast_llm ) - return OPEN_AI_CHAT_MODELS[llm_name] + return CHAT_MODELS[llm_name] @property def send_token_limit(self) -> int: diff --git a/autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py b/autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py index 53fadaa7c..ff08f4669 100644 --- a/autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py +++ b/autogpts/autogpt/autogpt/agents/prompt_strategies/one_shot.py @@ -122,7 +122,7 @@ class OneShotAgentPromptStrategy(PromptStrategy): 1. System prompt 3. `cycle_instruction` """ - system_prompt = self.build_system_prompt( + system_prompt, response_prefill = self.build_system_prompt( ai_profile=ai_profile, ai_directives=ai_directives, commands=commands, @@ -131,24 +131,34 @@ class OneShotAgentPromptStrategy(PromptStrategy): final_instruction_msg = ChatMessage.user(self.config.choose_action_instruction) - prompt = ChatPrompt( + return ChatPrompt( messages=[ ChatMessage.system(system_prompt), ChatMessage.user(f'"""{task}"""'), *messages, final_instruction_msg, ], + prefill_response=response_prefill, + functions=commands if self.config.use_functions_api else [], ) - return prompt - def build_system_prompt( self, ai_profile: AIProfile, ai_directives: AIDirectives, commands: list[CompletionModelFunction], include_os_info: bool, - ) -> str: + ) -> tuple[str, str]: + """ + Builds the system prompt. + + Returns: + str: The system prompt body + str: The desired start for the LLM's response; used to steer the output + """ + response_fmt_instruction, response_prefill = self.response_format_instruction( + self.config.use_functions_api + ) system_prompt_parts = ( self._generate_intro_prompt(ai_profile) + (self._generate_os_info() if include_os_info else []) @@ -169,16 +179,16 @@ class OneShotAgentPromptStrategy(PromptStrategy): " in the next message. Your job is to complete the task while following" " your directives as given above, and terminate when your task is done." ] - + [ - "## RESPONSE FORMAT\n" - + self.response_format_instruction(self.config.use_functions_api) - ] + + ["## RESPONSE FORMAT\n" + response_fmt_instruction] ) # Join non-empty parts together into paragraph format - return "\n\n".join(filter(None, system_prompt_parts)).strip("\n") + return ( + "\n\n".join(filter(None, system_prompt_parts)).strip("\n"), + response_prefill, + ) - def response_format_instruction(self, use_functions_api: bool) -> str: + def response_format_instruction(self, use_functions_api: bool) -> tuple[str, str]: response_schema = self.response_schema.copy(deep=True) if ( use_functions_api @@ -193,11 +203,15 @@ class OneShotAgentPromptStrategy(PromptStrategy): "\n", response_schema.to_typescript_object_interface(_RESPONSE_INTERFACE_NAME), ) + response_prefill = f'{{\n "{list(response_schema.properties.keys())[0]}":' return ( - f"YOU MUST ALWAYS RESPOND WITH A JSON OBJECT OF THE FOLLOWING TYPE:\n" - f"{response_format}" - + ("\n\nYOU MUST ALSO INVOKE A TOOL!" if use_functions_api else "") + ( + f"YOU MUST ALWAYS RESPOND WITH A JSON OBJECT OF THE FOLLOWING TYPE:\n" + f"{response_format}" + + ("\n\nYOU MUST ALSO INVOKE A TOOL!" if use_functions_api else "") + ), + response_prefill, ) def _generate_intro_prompt(self, ai_profile: AIProfile) -> list[str]: |