[FEATURE] Summarize the steps when max steps limit reached

[Hailong-am]

Is your feature request related to a problem?
Currently, when agent max step limit reached, it will return a default message like
Agent reached maximum iterations 20 without completing the task, It doesn't have too much meaning of what the agent have been done, it's better to summarize all the steps in this run and return the summary instead.

ml-commons/ml-algorithms/src/main/java/org/opensearch/ml/engine/algorithms/agent/MLChatAgentRunner.java

Lines 890 to 892 in be945ca

	String incompleteResponse = (lastThought.get() != null && !lastThought.get().isEmpty() && !"null".equals(lastThought.get()))
	? String.format("%s. Last thought: %s", String.format(MAX_ITERATIONS_MESSAGE, maxIterations), lastThought.get())
	: String.format(MAX_ITERATIONS_MESSAGE, maxIterations);

What solution would you like?

Add a option to let summarize all the steps when steps limit reached.

we have max_iteration for the total steps, maybe a new parameter like summarize_when_max_iteration: true

What alternatives have you considered?
A clear and concise description of any alternative solutions or features you've considered.

Do you have any additional context?
Add any other context or screenshots about the feature request here.

[pyek-bot]

hi @Hailong-am, thanks for opening the issue. A few comments:

1. If we do summarize, how do we inform the user that this is a summary after max_iteration? This is so that user knows next time they can increase max steps
2. Summary will be (max_steps + 1) LLM call after max_steps has reached. Do you think this would be issue if user assumes only x calls with be made based on their setting?

I think this parameter makes sense summarize_when_max_iteration: true

[Hailong-am]

1. If we do summarize, how do we inform the user that this is a summary after max_iteration? This is so that user knows next time they can increase max steps

we can return the final result as The max steps 20 reached, here is the summary of what did so far: ${summary_from_llm}, it tells the max steps reached and the summary.

3. Summary will be (max_steps + 1) LLM call after max_steps has reached. Do you think this would be issue if user assumes only x calls with be made based on their setting?>
   I think this parameter makes sense summarize_when_max_iteration: true

I think it's fine, as user have configured summarize_when_max_iteration as true, so user is aware that there will be an extra LLM call for the summarization.