Planning & Execution

Complex tasks have three big pain points — no structure, no control, and results that aren't directly usable — and the platform solves each with a dedicated capability.

Agent Todo Planning

Solves: complex tasks lacking structure.

• Planning node: upgraded from the tool selector node; with "Plan To-dos" enabled, it generates a task plan
• Agent node: with planning enabled, it automatically produces a To-do list and executes item by item according to the plan; at runtime you can watch the Agent progress through the plan
• Suited to multi-step, complex problems that need decomposition, making execution organized and trackable

Human in the Loop

Solves: fully automatic execution being uncontrollable.

Previously, an Agent in fully automatic mode ran straight to the end — mid-run deviations were only discovered after completion, making correction costly. Human in the Loop mode lets the AI proactively pause and confirm with the user during execution:

• Human confirmation at key steps: pause and wait for explicit authorization before high-risk operations — such as confirming investment risk level, modifying critical data, or tool calls involving privacy permissions — avoiding irreversible mistakes
• Controllable process: you can step in and adjust direction at any time during execution instead of only seeing the final result
• Higher trust: confirmation steps are embedded in the execution flow, closing the loop between human and machine judgment, and delivering more trustworthy results the first time in sensitive business scenarios

Recommendation for financial scenarios: for any flow involving write operations on account data or sending content externally, prefer HITL mode — consistent with the principle of Staying Within the Execution-Only Positioning that "AI does not make decisions for the customer."

Agent Answer Mode

Solves: answers that are merely "thought through" but not directly usable.

Agent answer mode organizes the answer around the goal and decomposes the task, rather than only outputting a linear chain of reasoning:

• Output form: task-oriented structures such as steps / plans / checklists, delivering directly actionable results with less rework
• Proactive completion: proactively fills in context and key tasks, so complex questions are resolved in one shot without multiple rounds of follow-up

Note: the "Deep Thinking mode" that previously coexisted with it has been retired; everything is now unified under Agent answer mode.

How the Three Relate

Agent answer mode  → determines "what the output looks like" (task-oriented structure)
Todo planning      → determines "how the process runs" (execute per checklist)
Human in the Loop  → determines "which steps need human sign-off" (pause at key points)

For complex and sensitive tasks, all three can be stacked: planned decomposition + confirmation at key steps + structured output.

Related Reading

• Agent Node Documentation — where the to-do and human-interaction switches are configured