thinking-makes-agents-introverted

name

thinking-makes-agents-introverted

description

Prevents the "introverted agent" problem where extended reasoning causes agents to give shorter, less informative responses in interactive/user-engaged tasks. Applies information disclosure (InfoDis) prompting to ensure agents proactively share state, enumerate options, and maintain transparency during multi-turn interactions. Use when: "help me build a customer service agent", "make my chatbot more helpful", "the agent isn't sharing enough information", "improve agent transparency", "fix terse agent responses", "design a user-facing assistant".

Thinking Makes Agents Introverted: Information Disclosure Prompting

This skill applies findings from Li et al. (2026) showing that mandatory reasoning (chain-of-thought, think-before-act patterns) causes LLM agents to become "introverted" — producing shorter responses, omitting critical information, and failing to enumerate options for users. The fix is Information Disclosure (InfoDis) prompting: explicitly instructing agents to proactively share observations, enumerate choices, and summarize system state at every user-facing turn. This technique improved performance across all seven models tested (GPT-5, Gemini-2.5-Pro, DeepSeek-V3.1, etc.) on customer service, booking, and phone assistant benchmarks.

When to Use

• When building or improving a multi-turn conversational agent (customer support, booking, admin assistant) that interacts with real users
• When an agent uses tool-calling or ReAct-style reasoning but produces terse, unhelpful responses to users
• When users complain the agent "doesn't tell me what's going on" or "just asks questions without explaining"
• When designing system prompts for agents that must balance internal reasoning with user communication
• When debugging agent failures where the root cause is information asymmetry — the agent knew something but didn't tell the user
• When an agent with chain-of-thought or <think> blocks performs worse on interactive tasks than expected

Key Technique

The Problem — Reasoning Induces Introversion. When agents are forced to reason before acting (via function-call thinking, ReAct prefixes, or <think> blocks), they systematically reduce the information they share with users. Empirically, DeepSeek-V3.1 dropped from 1.24 to 0.40 long responses (150+ tokens) per conversation trajectory when mandatory thinking was enabled (p<0.001). Gemini-2.5-Pro saw 37-38% performance degradation on customer service benchmarks. The mechanism: agents "spend" their communication budget on internal reasoning and then compress their user-facing output, omitting state summaries, option enumerations, and constraint disclosures.

The Fix — Information Disclosure (InfoDis) Prompting. The paper's core intervention is a system prompt addendum that explicitly instructs the agent to disclose information proactively. The key prompt: "You must interact with the user actively and disclose as much information as possible to ensure the user is well-informed about the current state and any potential changes." This is domain-agnostic — it attaches to any existing system prompt without task-specific examples. It reliably improved performance across all model families tested (+2.46% GPT-4o, +2.87% DeepSeek-V3.1, +1.30% Gemini-2.5-Pro average Pass@1).

The Taxonomy — Two Disclosure Types. The paper identifies two atomic categories of information disclosure that agents should produce: (1) Observation Summarization — retelling the current state, prior observations, and planned next steps; (2) Choice Enumeration — listing available options in structured form so the user can make informed decisions. Agents that produce both types at decision points outperform those that skip either.

Step-by-Step Workflow

1. Identify the interaction pattern. Determine whether the agent is user-engaged (multi-turn, user makes decisions mid-conversation) vs. autonomous (agent completes task alone). This technique applies specifically to user-engaged agents where information exchange matters.

2. Audit the existing system prompt for introversion risk. Check if the prompt includes mandatory reasoning instructions (e.g., "think step by step before responding", "always use the think tool first", ReAct format requirements). These are introversion triggers.

3. Append the InfoDis addendum to the system prompt. Add a disclosure directive after any reasoning instructions. The directive must be explicit and imperative:

   INFORMATION DISCLOSURE REQUIREMENT: You must interact with the user actively
   and disclose as much information as possible. At every user-facing response:
   - Summarize the current state of relevant data/systems you have observed
   - Enumerate available options or choices the user can make
   - Disclose any constraints, limitations, or edge cases that affect the user's decision
   - Explain what you plan to do next and why
   

4. Ensure disclosure survives reasoning blocks. If the agent uses <think> tags or a think tool, add an explicit instruction that the user-facing response AFTER thinking must contain full disclosure — reasoning output must not substitute for user communication.

5. Implement the two-type disclosure check. For each agent response that precedes a user decision, verify it contains: (a) an observation summary of current state, AND (b) an enumeration of available choices. If either is missing, the response risks causing downstream failure.

6. Test with failure propagation scenarios. Create test cases where the agent must disclose a non-obvious constraint (e.g., "individual items in a multi-item order cannot be cancelled separately"). Run with and without InfoDis to verify the agent surfaces the constraint.

7. Measure response informativeness. Track average response token count and count of atomic information units per trajectory. A drop in either metric after adding reasoning instructions signals introversion — apply stronger disclosure prompting.

8. Balance disclosure with conciseness. InfoDis should not produce walls of text. The goal is to disclose decision-relevant information, not narrate every internal step. Frame the directive around what the user needs to know to make their next decision.

9. Iterate across domains. The InfoDis prompt is domain-agnostic, but the specific constraints and options to disclose are domain-specific. For each new domain, identify the top 3-5 information types that users need (e.g., for retail: return policies, order status, refund timelines; for booking: availability, pricing tiers, cancellation rules).

Concrete Examples

Example 1: Customer Service Agent — Order Cancellation

User: "I want to cancel one item from my order."

Without InfoDis (introverted response):

I'll look into that for you. Which item would you like to cancel?

With InfoDis (transparent response):

I've pulled up your order #4521. It contains 3 items:
1. Wireless Headphones ($89.99) - Ships March 5
2. USB-C Cable ($12.99) - Ships March 3
3. Phone Case ($24.99) - Ships March 3

Important: This order is processed as a single shipment. Our system only
supports cancelling the entire order, not individual items. Here are your options:

- Cancel the full order and reorder the items you want to keep
- Keep the order as-is and return the unwanted item after delivery (free returns within 30 days)
- I can check if the warehouse can split the shipment (not always possible)

Which would you prefer?

Why it matters: The introverted response leads the user to name an item, only to be told after that individual cancellation isn't possible — wasting a turn and frustrating the user. The transparent response surfaces the constraint upfront and enumerates actionable alternatives.

---

Example 2: System Prompt Design for a Booking Assistant

User: "Write me a system prompt for a flight booking agent that uses tool calling."

Approach:

1. Include standard role and tool-calling instructions
2. Add reasoning capability without making it suppress disclosure
3. Append the InfoDis addendum targeting booking-specific information

Output:

You are a flight booking assistant with access to the following tools:
[search_flights, book_flight, cancel_booking, get_booking_details, check_seat_availability]

When handling a request, you may use internal reasoning to plan your tool calls.
However, your responses to the user must ALWAYS include:

INFORMATION DISCLOSURE RULES:
1. After every search, summarize results: show flight options with prices,
   times, layovers, and seat availability in a structured format
2. Before any booking action, confirm: total cost, fare class restrictions
   (change fees, cancellation policy, baggage), and payment method
3. When constraints exist (sold-out flights, fare class restrictions,
   schedule conflicts), disclose them BEFORE asking the user to choose
4. Enumerate alternatives when the user's first choice is unavailable —
   never just say "that's not available" without offering next steps

Priority: User-facing transparency is MORE important than internal reasoning
efficiency. A well-informed user who makes one correct decision is better
than a confused user who requires three clarification rounds.

---

Example 3: Debugging a Terse Agent

User: "My ReAct agent gives great reasoning in its Thought blocks but its responses to users are just one-liners. How do I fix this?"

Approach:

1. Diagnose: This is the classic introversion pattern — reasoning consumes the "communication budget"
2. Separate the reasoning channel from the disclosure channel
3. Add explicit post-reasoning disclosure instructions

Output:

# The problem: ReAct format makes agents compress user responses
# Original prompt fragment:
REACT_PROMPT = """At each step, generate:
Thought: <your reasoning>
Action: <tool call or response>"""

# The fix: Add a Disclosure step between Thought and user-facing Action
REACT_WITH_INFODIS = """At each step, generate:
Thought: <your internal reasoning — not shown to user>
Disclosure Plan: <list what the user needs to know before their next decision>
Action: <tool call OR user response that includes all items from Disclosure Plan>

CRITICAL: Your Action response to the user must be self-contained and informative.
The user cannot see your Thought block. Everything decision-relevant must appear
in the Action response. Specifically:
- Summarize what you observed from any tool calls
- List available options with key differentiators
- State any constraints or limitations that affect the user's choice
- Propose a recommended next step with reasoning the user can evaluate"""

Best Practices

Do:

• Always append InfoDis instructions AFTER any reasoning/thinking instructions in system prompts, so disclosure is the final imperative the model processes
• Use the two-type framework: every pre-decision response should contain both a state summary and an option enumeration
• Test for introversion by measuring response token counts with and without reasoning enabled — a statistically significant drop indicates the problem
• Frame disclosure around the user's next decision, not the agent's internal process

Avoid:

• Don't assume chain-of-thought or <think> blocks will make agents more helpful to users — the research shows the opposite in interactive settings
• Don't conflate internal reasoning quality with external communication quality — an agent can reason perfectly but still fail by not telling the user what it found
• Don't add reasoning mandates ("you MUST think before every response") without a corresponding disclosure mandate — this is the most common cause of the introversion pattern
• Don't over-disclose: dumping raw tool outputs or internal state is not the same as curated, decision-relevant transparency
• Don't use InfoDis for autonomous agents that don't interact with users mid-task — it's specifically designed for user-engaged multi-turn scenarios

Error Handling

Failure Mode	Symptom	Fix
Constraint omission	Agent asks user to choose, then rejects their choice due to undisclosed rule	Add domain-specific constraints to the InfoDis prompt as explicit disclosure targets
Option truncation	Agent shows 1-2 options when 5+ exist	Add "enumerate ALL available options" to disclosure rules; set a minimum count if applicable
State staleness	Agent summarizes outdated state after multiple tool calls	Require re-summarization after every state-changing tool call, not just at conversation start
Disclosure overload	Responses become too long and users disengage	Prioritize disclosure by decision-relevance; use structured formatting (tables, bullet points) to increase density
Reasoning leak	Agent's <think> content bleeds into response but replaces actual disclosure	Ensure the prompt separates reasoning output from user-facing output with distinct formatting rules

Limitations

• Autonomous tasks: InfoDis is designed for user-engaged agents. For fully autonomous pipelines where no human is in the loop, standard reasoning techniques remain effective.
• Token cost: Transparent responses are longer. In latency-sensitive or cost-constrained deployments, the added tokens from disclosure must be weighed against the reduced back-and-forth from better-informed users.
• Domain knowledge required: The InfoDis prompt is domain-agnostic, but effective disclosure requires knowing what to disclose. In novel domains, you may need to iteratively discover which information types users need most.
• Measured improvements are moderate: The paper reports +1.3% to +2.87% average Pass@1 improvements with InfoDis. The technique prevents degradation more than it creates breakthroughs — its primary value is avoiding the introversion penalty of reasoning.
• Not a substitute for good tools: If an agent lacks access to the right tools or data, no amount of disclosure prompting will help. InfoDis improves communication, not capability.

Reference

Li, J., Oh, C., Choi, H. K., Wang, J., & Li, S. (2026). Thinking Makes LLM Agents Introverted: How Mandatory Thinking Can Backfire in User-Engaged Agents. arXiv:2602.07796v1. https://arxiv.org/abs/2602.07796v1

Key sections to read: Section 4 (Response Taxonomy Analysis) for the two-type disclosure framework; Section 5 (InfoDis Prompting) for the exact intervention; Table 2 for per-model degradation numbers; Figure 4 for failure propagation case studies showing how omitted information cascades into task failure.