When processing the response of an LLM during a task, it should be possible to return as a task response a list of next queries and which agent should be executing the query. This would allow for example the ability to run multiple personas/roles as agents and have them process requests.

This can already be done through the HTTP server endpoint but is untested (may have race conditions).

Allow the operator/user to visualize the current stack of tasks being worked on by the agent.

Allow calling echo "Some query" | python cli.py, which should behave just like python cli.py "Some query".

Only allow authenticated users and agents to call HTTP endpoints.

Allow users to insert additional prompts before and after the query generated by a given task.

For example
post_prompt=Be concise.

could be added to a query-multiple-personas task.

Introduce the concept of worker and reviewer.

Share a global budget among many agents.

An initial implementation could be to give a part of the current agent's remaining budget to the new agent. Once the agent terminates, it could report how much budget it used itself. Alternatively, we could record how much budget was spent, and at regular interval adjust the remaining budget.

A more "elegant" solution would be to use a shared database or memory, possibly using dask Lock.

Have a table with the original query, the task, the current state.
Have a loop that query notion to get any entry that isn't started yet and process it.

Design a prompt used to generate python code which can solve the specified goal.
The python code can be a skeleton and get more detailed over time.

In the QueryMultiplePersonas task, it should be possible to call the task many times and not necessitate any call to the LLM if all ther personas have already been resolved.

Select which personas to query when using QueryMultiplePersonas or SummarizeMultiplePersonas using different methods, such as querying the LLM for its opinion, picking randomly, etc.

Run AutoGPT as an HTTP server that can process queries in the same format as the CLI (i.e. task + request/query).
Return in JSON the response and all associated information (e.g., consumed token, cost).

Ideally the HTTP server framework should support HTTP/2 to reduce the need to reconnect between client and server.

When launching the CLI, it should be possible to specify a sequence of tasks to execute sequentially (maybe a graph of tasks?).

For instance, we currently have SummarizeMultiplePersonas as a task, which is in itself a collection of calls to Simple, followed by a SummarizeResponses.

In #36 we want to implement a personas selector. The idea would be to run SelectPersonas, then QueryMultiplePersonas with the returned list, and finally SummarizeResponses.

We should be able to get rid of SummarizeMultiplePersonas and replace it with QueryMultiplePersonas,SummarizeResponses.

The goal is to increase reusability while reducing the need for tasks that are composite of other tasks.

Find a ChatGPT web UI on GitHub and adapt it to call the AutoGPT HTTP backend.

This might be achievable by using a library that already offers support for multiple backends, e.g., LangChain.

• Claude (Anthropic)
• Coral (Cohere)
• Bard (Google)
• Replicate

We currently support a monetary budget to determine how long the agent should run.
We should be able to also specify a time constraint, after which the agent will stop.

Some complex queries could be submitted as a graph, where an initial query is submitted, which can result in many agents being spawned, after which we would like to take the generated outputs and feed it back into a single query (e.g., submit a single query to multiple roles, then take their individual feedback and summarize them).

Make it possible for multiple personas to have a discussion between each other. The number of personas can be unlimited.

Ideally we would spin up one agent per persona and have the messages go back and forth between the agents.
All the conversations should be part of a unique session for each agent.
If an agent budget is consumed, they will stop participating in the discussion.

Implement an endpoint compatible with https://platform.openai.com/docs/api-reference/chat/create such that it is easy to replace https://api.openai.com/v1/chat/completions with https://my-server/v1/chat/completions.