## Objectives π―
Nichebiche opened this issue Β· comments
Objectives π―
- AutoGPT should be able to:
- Handle tasks of arbitrary complexity
- Recover and continue in case of unexpected intermediary results, e.g. tool failure or unavailability of a non-essential 3rd party service
- AutoGPT should be optimally productive in terms of achieved results per $ of LLM expense
SWOT π§
Strengths πͺ
- Exceptions and other problems during tool execution are handled gracefully and fed back to the LLM
Weaknesses π§
- Linear planning & stepping
- Limited historical context, even with compression of the
ActionHistory
in the prompt - Failed execution branches remain in historical memory and take up space
- Limited historical context, even with compression of the
Opportunities π
- Non-linear planning strategies could provide a combination of better (relevant) context retention and recovery on failure
- Nested planning strategies could allow handling tasks of arbitrary complexity
Threats β οΈ
- Nested planning carries a risk of (semi-)infinite recursion
- Context retention could become less relevant as models with lower token cost and larger effective context windows become available
- Nested planning could be the kryptonite that enables AutoGPT to do something scary
Related πͺ
Actionables
- #4107
- Implement execution of multiple actions at once
- OpenAI supports this natively in their new GPT models since 2023-11-06
- Implement action piping/chaining (maybe through code bindings/API?)
- Implement intent β implement β execute β evaluate flow (instead of handling everything except execution in the same one-shot prompt)
- The βimplementβ step would be a good place to integrate Procedural Memory.
Notes
- The terms "actions", "commands", "tools" are used interchangeably and refer to the same concept.
As mentioned in the post, this is a duplicate of #6964. If you have something to add to the roadmap item, please comment below it.