# Deep Research and Deep Thinking

Give your AVCodex agent the ability to run multi-source investigations and show step-by-step reasoning. Useful when a design engineer is comparing matrix switchers across Crestron, Extron, and AMX, when a sales engineer is researching a new vertical (broadcast, command and control, houses of worship), or when a programmer needs the agent to actually think through a control-system design instead of guessing.

---

AVCodex agents support two advanced reasoning capabilities beyond standard chat: **Deep Research** for multi-source investigation and **Deep Thinking** for visible, step-by-step reasoning. Both produce more thorough, transparent answers to complex questions.

> **Warning:** Deep Research and Deep Thinking require a Builder plan or higher. Not available on the Free tier.

## Deep Research

Deep Research spawns multiple parallel research agents that independently search the web, read pages, and compile findings. Results are synthesized into a single cited report.

### When the Agent Uses Deep Research

The agent automatically reaches for Deep Research on questions that need broad investigation:

- Manufacturer comparisons across multiple lines (Crestron NVX vs Extron NAV vs AMX SVSI for a 40-room AVoIP refresh, for example).
- Technical deep dives needing several authoritative sources.
- "Compare X vs Y vs Z" questions on multiple dimensions.
- Industry reports and competitive analyses.
- Multi-faceted questions where a single web search wouldn't be enough (specifying a DSP for a complex broadcast workflow, for instance).

It does **not** use Deep Research for simple factual questions, casual conversation, follow-ups where context already exists, or questions about the user's own data.

### How It Works

**1. Decomposition.**

The agent breaks the question into 2-6 focused sub-questions, each targeting a distinct aspect. "Compare Crestron NVX, Extron NAV, and AMX SVSI for enterprise AVoIP" might become sub-questions about latency and PTP support, codec efficiency, control system integration, hardware cost per endpoint, and product roadmap.

**2. Round 1: Parallel Investigation.**

Each sub-question gets its own research agent that independently searches the web, reads full pages, and follows links to primary sources (manufacturer datasheets, Avixa whitepapers, integrator forums). These agents run in parallel for speed.

**3. Gap Analysis.**

After Round 1, an analysis step reviews findings for gaps, contradictions, and areas needing more depth. If the research is already comprehensive, it moves straight to synthesis.

**4. Round 2: Follow-Up Research (if needed).**

Up to 4 follow-up agents investigate gaps. They have context from Round 1, so they can verify claims, cross-reference sources, and fill in missing information.

**5. Synthesis.**

All findings from both rounds are synthesized into a single, structured report organized by theme. Every claim includes inline citations with links to sources.

### What the User Sees

During Deep Research, the chat UI shows real-time progress:

1. **Plan**: the sub-questions being investigated.
2. **Research progress**: which agents are running, how many sources have been found.
3. **Gap analysis**: whether follow-up research is needed.
4. **Final report**: a comprehensive response with inline source citations.

A typical Deep Research operation takes 30-90 seconds depending on the complexity and the number of follow-ups.

### Example Prompts

- "Research the latest PTP-aware AVoIP product lines from Crestron, Extron, and AMX. Include latency, switch requirements, and codec details."
- "Compare the top five conference room cameras for a hybrid 12-person huddle space. Cover Logitech, Poly, Yealink, Sony, and Cisco."
- "What's the current state of Dante Domain Manager 1.5 in production deployments? Pull from real integrator field reports."
- "Analyze the competitive landscape for cloud-managed digital signage platforms targeting K-12."

## Deep Thinking

Deep Thinking enables visible, structured reasoning before the agent answers. Instead of jumping straight to a response, the agent creates a step-by-step plan and works through it, showing the reasoning process.

### When the Agent Uses Deep Thinking

The agent automatically uses Deep Thinking for questions that benefit from structured analysis:

- Comparing two or more options, approaches, or strategies (Q-SYS Lua versus a Crestron SIMPL# scheduling module, for example).
- Questions with multiple distinct parts that each need analysis.
- Providing recommendations based on specific criteria (a DSP recommendation given budget, channel count, and Dante requirements).
- Complex analysis requiring structured reasoning.
- Cross-referencing multiple knowledge base sources (Crestron, Biamp, and Extron datasheets all together).
- Combining web research with internal knowledge base data.

It does **not** use Deep Thinking for simple factual questions with a single clear answer, casual conversation, brief follow-ups, or yes/no questions.

### How It Works

**1. Plan Creation.**

The agent creates a reasoning plan with 2-8 steps. Each step maps to a specific action: searching knowledge, browsing the web, analyzing data, comparing options, or synthesizing conclusions.

**2. Step Execution.**

The agent works through each step in order, using the right tools (knowledge search, web browse, code execution) as needed.

**3. Comprehensive Response.**

After completing all steps, the agent produces a thorough response informed by the structured reasoning.

### What the User Sees

During Deep Thinking, the chat UI shows the reasoning steps as the agent works through them. Each step has a label ("Searching knowledge base for QSC Core 110f I/O specs," "Comparing DSP block costs," "Synthesizing recommendation") and updates in real time.

This transparency helps users understand how the agent reached its answer. Important when an AV programmer or design engineer needs to defend a specification choice to a client or a procurement team.

### Example Prompts

- "Should we use a Q-SYS Core 510i or a BiampVocia for this 1,200-seat house of worship? Consider channel count, paging requirements, and integrator familiarity."
- "Review our standard 12-person huddle room scope and identify the three biggest risks for a fixed-bid RFP."
- "What's the best programming approach for cross-room scheduling when half the rooms are Crestron and half are Q-SYS?"
- "Analyze our Q3 truck-roll data and recommend focus areas for Q4 preventive maintenance."

## Deep Research vs Deep Thinking

| | Deep Research | Deep Thinking |
|---|---|---|
| **Purpose** | Gather information from many external sources | Structure reasoning over available information |
| **Speed** | 30-90 seconds | 10-30 seconds |
| **Sources** | Parallel web search agents | Knowledge base, web, code execution |
| **Output** | Cited research report | Structured analysis with visible reasoning |
| **Best for** | "What's the current state of X?" | "What should I do about X?" |
| **Agents** | 2-6 parallel sub-agents plus follow-up agents | Single agent with multi-step plan |

In practice, the agent often combines both: Deep Research to gather information, Deep Thinking to analyze it. The agent decides which capability to use (or both) based on the question.

## Tier Requirements

Both capabilities require a **Builder** plan or higher:

| Tier | Deep Thinking | Deep Research |
|------|---------------|---------------|
| Free | No | No |
| Builder | Yes | Yes |
| Studio | Yes | Yes |
| Studio Pro | Yes | Yes |
| Enterprise | Yes | Yes |

## Billing

Both Deep Research and Deep Thinking consume tokens through your organization's Stripe Token Billing balance. Deep Research is more expensive because it runs multiple parallel LLM calls (one per sub-question, plus gap analysis and synthesis). A typical Deep Research operation uses 5-15x the tokens of a standard chat response.

Deep Thinking has modest overhead since it uses a single model with structured planning. Typically 2-3x a standard response.

> **Tip:** If your agent mostly serves simple Q&A (a quick room-support agent answering "How do I start the projector?", say), users will rarely trigger these capabilities. The agent only activates them for questions that benefit from structured reasoning or multi-source research.

*AVCodex · Your AV expertise. Amplified by AI.*