muellerberndt / darwin-gpt Goto Github PK

This ticket should be tackled after #3 has been implemented.

Replace the WRITE_EXEC function with a new function that allows the agent to replace itself with a new version (instead of blindly forking additional processes). Specs:

1. The agent should attempt to run the new version of the code received by GPT in a new process and capture STDOUT/STDERR.

2.a. If the new code runs without error, the agent should terminate.
2.b. In case of an error or crash, the agent should send an error report to GPT and request fixed code, and proceed with step 1.

The number of retries should be configurable.

Desired agent behavior

With this ticket, the agent should gain the capability to reliably update itself with new functionality, in particular new commands. Make sure that the prompt makes this clear to the agent and test whether GPT understands the process correctly.

Add a command that allows the agent to retrieve & summarize a website and store it in memory.

Add the necessary types of memory to the agent so that it can maintain context over multiple commands. The implementation should be as simple as possible (use local files, no external APIs). Keep in mind that the code needs to remain compact.

Add a command that allows the agent to retrieve search results from Google.

Instead of only writing and executing code once, create a loop that allows the agent to execute multiple commands (as received by GPT) that map to Python functions. To do this, we should create a standard interface (it's probably best to use a JSON format as Auto-GPT - investigate this). The loop should work as follows:

1. The agent prompts GPT for the next command;
2. GPT responded with the desired command and the command arguments;
3. Agent executes the command and sends the command output, along with a query for the next command, to GPT (repeat in a loop).

The JSON request could look as follows (example - feel free to improve on this):

{"command": {
  "name:" "write_and_execute",
  "args": [
     "code": "(escaped_code)",
     "filename": "agent.py"
}}

Note that this is similar to how Auto-GPT and babyAGI work, however, we have to keep the code as compact as possible and in a single file to allow for easy self-replication.

For this ticket, implement only a single command named WRITE_EXEC(code) or similar that writes code to a file, forks a new process, and executes the updated code (exactly like it is currently done).