The ability to chain multiple commands or actions together could significantly increase the efficiency of AutoGPT.

Related:

• Command execute_command_chain returned: Unknown command 'execute_command_chain'. Please refer to the 'COMMANDS' list for available commands and only respond in the specified JSON format. #949
• Multi-function calling
• Persistent Python session / Jupyter Notebook integration #5132

Notes & ideas for implementation

• Execution MUST stop when a command in the chain fails. In case of an execution tree, the branches other than the failed branch can continue executing.

The simplest way?

The simplest way is probably to amend the current command execution interface to support calling multiple functions which are executed in order, and of which the output can be used as arguments for the next command:

"thoughts": "[thoughts]",
"commands": [
  {
    "name": "get_text_from_webpage",
    "args": {
      "url": "https://en.wikipedia.org/wiki/Otters"
    },
    "outputs": {
      "content": "$otters_content"
    }
  },
  {
    "name": "write_file",
    "args": {
      "content": "$otters_content"
    }
  }
]

Execution tree

More complex, harder to implement and probably more error prone, but worth considering:

"thoughts": "[thoughts]",
"commands": [
  {
    "name": "get_text_from_webpage",
    "args": {
      "url": "https://en.wikipedia.org/wiki/Otters"
    },
    "outputs": {
      "content": "$otters_content",
    },
    "pipe_into": [
      {
        "name": "write_file",
        "args": {
          "content": "$otters_content"
        }
      }
    ]
  }
]

Command interface -> script-like execution

We could profit from totally rethinking the command execution interface. For example, if we implement #6253 + #5132, and we exposed all available commands as functions that can be imported or executed in that environment, it could be as simple as this:

• User: Please scrape the content of https://en.wikipedia.org/wiki/Otters and write it to a text file```

• AutoGPT:

  <execute>
  content = get_text_from_webpage("https://en.wikipedia.org/wiki/Otters")
  write_file("otters.txt", content)
  </execute>
  

  I have written the content of the webpage to the file otters.txt!

Currently, it would be more like:

• User: Please scrape the content of https://en.wikipedia.org/wiki/Otters and write it to a text file```

• AutoGPT: I will now get the content from the webpage.
  Executing get_text_from_webpage("https://en.wikipedia.org/wiki/Otters")

• AutoGPT: Now that I have scraped the webpage, I will save the content to a file called otters.txt.
  Executing write_file("otters.txt", "[lots of text/tokens here, very expensive and possibly slow!]")

This is both slower and significantly more expensive, because we're copy-pasting data using an LLM.