[Question]: Agent time and date understanding and translating relative dates to tools date ranges

[cesarqdt]

Question Validation

• I have searched both the documentation and discord for an answer.

Question

My question is this one: How to give agents temporal context for relative date queries?

When I ask my agent "what happened 2 weeks ago?" or "show me today's data," it needs to understand the current date and translate that into concrete date ranges for the tools it calls. Right now, the agent struggles with this kind of relative temporal language.
I'm building financial analysis agents that handle queries with dates like "last month," "today," or "30 days ago," and the agent should be smart enough to convert these into actual date ranges it can pass to its tools. Instead, it either fails to make the right tool calls or passes the responsibility back to the user.

What's the best way to give an agent this temporal understanding in LlamaIndex? Should I inject the current date into the system prompt - and if so, how should I implement that properly? - and create a dedicated temporal tool, or is there a pattern I'm missing?

Thank you!

[logan-markewich]

Put the current date and time into the system prompt. You can just use datetime and throw this into the system prompt directly no?

agent = FunctionAgent(..., system_prompt="... The current date is:...")

[cesarqdt]

Thanks! we've tried this, but doesn't always get it right. So we have another agent that is responsible for pulling data from an API and that agent needs a start_date and end_date. Sometimes if the user says I want the data from 2 years ago until today, the agent does not infer the right date ranges and sets the start_date to a random date in the past and same for today's date.

We have an orchestrator that should route correctly but it does not always work. So for example:

date_agent = FunctionAgent(
    name = "DateAgent",
    description = "Useful for understanding dates",
    tools = [get_now_tool],
    llm = get_llm(),
    system_prompt = "The current date is {datetime.now(timezone.utc}",
    can_handoff_to = ["SearchAgent", "FinanceAgent"],
)

orchestrator_agent_workflow = AgentWorkflow(
    agents = [ 
        date_agent,
        search_agent,
        finance_agent,
    ],
    root_agent = search_agent.name,
)


orchestrator_agent_workflow.run("Get me data for this company from last year to today")
>>> Response
I currently do not have access to historical price data

For the query "Get me data for this company from last year to today", the agent would infer the data without sending to the date_agent or would simply call the date_agent and then skip the finance_agent and return something like "I don't have data for that range" but that's not true, because we know that the data exists for those ranges. So what's the right approach to allow the agent to understand and parse dates properly and hand them out correctly as the parameters for the tools of the other specialized agents?