Thanks — I hear you: your agent is calling

get_instructions_tool

and then immediately

transfer_to_live_agent

in the same turn, skipping the required data-collection steps (call reason / name / phone). I’ll walk through why that happens and give actionable fixes you can apply. Why this happens • The model can decide to call multiple tools in a single turn when the LLM integration allows parallel tool calls. If

parallel_tool_calls

is enabled, the model may try to invoke both tools concurrently instead of following your intended sequence. • LLMs sometimes attempt a best-effort handoff if a transfer tool is available, so without protections the model can jump to

transfer_to_live_agent

even when required fields are missing. Practical fixes you can apply 1. Disable parallel tool calls (if configurable) ◦ Turn off

parallel_tool_calls

for your OpenAI/Gemini integration so the model is constrained to one tool call per turn. This prevents simultaneous invocations and makes ordering more predictable. ◦ If you cannot disable it globally, disable it for the agent profile or request pipeline that handles handoffs. 1. Make

transfer_to_live_agent

require preconditions ◦ Have the

transfer_to_live_agent

tool check session state (

userdata

or

RunContext

) before performing the transfer. ◦ If required fields (call reason, name, phone) are missing, return a structured response that: ▪︎ Explains which fields are missing, and ▪︎ Prompts the user to provide them (do not perform the transfer). ◦ Only execute the actual transfer when all required data are present. 1. Track collection with session state ◦ Use

userdata

(or the session context mechanism you already use) to store flags like

has_name

,

has_phone

,

call_reason_collected

. ◦ Tools and the agent’s instruction template should read/write these flags so both model and tools share the same truth about progress. 1. Encapsulate collection in a Task ◦ Create a

CollectUserInfoTask

or similar

AgentTask

that owns the flow of asking for and validating name/phone/reason. A task takes control of the session until its goal is achieved, which forces step-by-step completion before the handoff is allowed. ◦ Only allow the

transfer_to_live_agent

to be callable after the task marks completion. 1. Make tool responses explicit and structured ◦ Return deterministic, short outputs from

get_instructions_tool

(e.g., JSON or a labeled response) so the LLM can’t infer that a transfer should happen immediately. ◦ Example:

get_instructions_tool

returns {"instructions_collected": true, "next":"ask_for_name"} — this reduces ambiguity and improves the model’s decision-making. Quick checklist to implement 1. Check and disable

parallel_tool_calls

in your LLM plugin/configuration. 2. Add precondition checks inside

transfer_to_live_agent

. 3. Persist collected fields in session state and read them before transfer. 4. Optionally move the collection to an

AgentTask

that must complete first. 5. Make tools return structured outputs to reduce ambiguity. If you’d like, share: • The tool definitions (code for

get_instructions_tool

and

transfer_to_live_agent

), and • The agent prompt/instruction template or relevant config (including whether

parallel_tool_calls

is enabled). With those I can point out exactly where to add the checks or provide a concrete code example for the transfer precondition and task implementation. Sources: Together AI LLM plugin guide | LiveKit Documentation | Agents and handoffs | LiveKit Documentation | Tasks and task groups | LiveKit Documentation