5K

Verbose Reasoning Transparency

Lets users open a verbose view on demand that shows the agent's interpretation, tool choices, intermediate steps, and raw tool outputs

best practice Updated 2026-10-13 By Nikola Balic (@nibzard) Based on Boris Cherny (via Claude Code)
View .md Edit on GitHub
Add to Pack
or

Saved locally in this browser for now.

Cite This Pattern
APA 
Nikola Balic (@nibzard) (2026). Verbose Reasoning Transparency. In *Awesome Agentic Patterns*. Retrieved October 14, 2026, from https://agentic-patterns.com/patterns/verbose-reasoning-transparency
BibTeX 
@misc{agentic_patterns_verbose-reasoning-transparency,
  title = {Verbose Reasoning Transparency},
  author = {Nikola Balic (@nibzard)},
  year = {2026},
  howpublished = {\url{https://agentic-patterns.com/patterns/verbose-reasoning-transparency}},
  note = {Awesome Agentic Patterns}
}

Use when

  • Users must debug unexpected agent output
  • Users need to learn how to prompt the agent better
  • High-stakes tasks where users must know why the agent acted

Avoid when

  • Verbose output would expose sensitive system prompts or credentials
  • Token overhead is not acceptable

Problem

AI agents, especially those using complex models or multiple tools, can sometimes behave like "black boxes." Users may not understand why an agent made a particular decision, chose a specific tool, or generated a certain output. This lack of transparency can hinder debugging, trust, and the ability to effectively guide the agent.

Solution

Implement a feature that allows users to inspect the agent's internal "thought process" or reasoning steps on demand. In Claude Code, Ctrl+O toggles the transcript viewer, which shows detailed tool usage and execution per assistant message; the --verbose flag (and the matching verbose setting) raises output detail for a session.

When activated, the verbose output might reveal:

  • The agent's interpretation of the user's prompt.
  • Alternative actions or tools it considered.
  • The specific tool(s) it selected and why (if available).
  • Intermediate steps or sub-tasks it performed.
  • Confidence scores or internal states, where a harness exposes them (Claude Code does not).
  • Raw outputs from tools before they are processed or summarized.

This transparency helps users understand the agent's decision-making process, identify issues if the agent is stuck or producing incorrect results, and learn how to prompt more effectively.

Faithfulness caveat: displayed reasoning is not guaranteed to be what actually drove the answer. Anthropic finds that reasoning models often do not verbalize their real reasoning (Reasoning models don't say they think), and OpenAI reports that models under monitoring can learn to obfuscate (arXiv:2503.11926). Use verbose views to debug behavior; do not treat them as proof of the model's intent.

Distinct from Chain-of-Thought Monitoring & Interruption, which watches the reasoning stream continuously for safety: this pattern is a human opening the view on demand to debug or learn.

Example (transparency activation)

sequenceDiagram participant User participant Agent participant UI as Interface User->>Agent: Complex task request Agent->>Agent: Process internally Agent-->>User: Standard output User->>UI: Ctrl+O (transcript viewer) UI->>Agent: Request verbose details Agent-->>UI: Internal reasoning steps Agent-->>UI: Tool selection rationale Agent-->>UI: Raw tool outputs UI-->>User: Detailed transparency view

How to use it

  • Debugging agents that produce incorrect or unexpected outputs
  • Learning how to prompt more effectively by studying agent reasoning patterns
  • Building trust in high-stakes scenarios where understanding "why" matters
  • Complementing human-in-the-loop approval workflows with transparency

Trade-offs

  • Pros: Enables debugging of unexpected agent behavior, supports prompt engineering, and builds trust through explainability.
  • Cons: Requires careful handling of sensitive information (system prompts, credentials); displayed reasoning can be unfaithful to the model's actual computation (see the caveat above), so it can mislead as well as explain.

References

Source