How to Write Advanced Prompts for Search Agents in 2026

How to write effective prompts for modern search agents

1. Know your tools: Exa’s Agent Ultra

Agent Ultra represents the pinnacle of Exa’s Agent API. It coordinates thousands of subdomains and sub-agents to create extensive, high-quality datasets. To harness its capabilities, the prompt must clearly define the objective, the type of sources, and the desired output format.

Example prompt for Agent Ultra:

Objective: Create a list of 50 startups using AI for sustainability.
Format: List with name, sector, and short description (max 150 characters per entry).
Restrictions: Use only sources published in 2026, include URLs and contact details where available.
Steps: 1) Identify primary sources (industry magazines, market reports). 2) Delegate data collection to a sub-agent. 3) Collect, verify, and compile data.
Status: Pending initiation.

2. When to use MSEB for sound processing

Google Research’s Massive Sound Embedding Benchmark (MSEB) is the industry standard for custom audio encoders. Whether you’re classifying sound clips, clustering them, or extracting segments, a well-structured prompt can guide the model to the optimal configuration.

Example prompt for MSEB:

Task: Implement a compact audio encoder for 10-class classification (e.g., engines, speech, music).
Benchmark: Use MSEB v2 contract (256-token embedding, contrastive loss).
Metrics: Top-1 accuracy > 85% on validation set.
Instructions: 1) Load dataset into tf.data. 2) Define a model with a ConvNeXt-based encoder. 3) Train for 30 epochs with lr=1e-4.
Output: Save checkpoint as sound_encoder_mseb.ckpt.

3. Leverage Julia 1 for decision-making tasks

Julia 1, the open-source 144.3 million-parameter decision model released by Supersonic Labs, excels in scenarios requiring the evaluation of multiple options based on provided context. Its architecture is designed for use via chat-style prompts.

Example prompt for Julia 1:

Context: A user has a budget of €5,000 and wants a laptop for 3D graphics.
Options:
1) Dell XPS 15 (CPU i7, GPU RTX 4060)
2) ASUS ProArt Studiobook (CPU i9, GPU RTX 4070)
3) Lenovo ThinkPad P1 (CPU i9, GPU RTX 4060)
Question: Which option offers the best value for money? Justify your answer.
Formatting: List the choice and reason in ≤ 100 words.

4. Advanced prompt engineering techniques

To achieve consistent results with these agents, combine proven methods:

• Chain-of-thought prompting:Ask the agent to break the problem into logical steps.
• Few-shot learning:Provide concrete examples within the prompt to clarify tone and structure.
• Dynamic prompts:Use variables and functions (e.g., {{data_source}}) to make the prompt adaptable.
• Budget control:Specify the maximum number of tokens to avoid overloads.
• Feedback-based iteration:Collect results, analyze errors, and refine the prompt accordingly.

5. Practical workflow: From query to result

Follow this schema to transform a business request into an effective prompt:

1. Define the objective:What is the desired end result?
2. Select the tool:Agent Ultra for extensive research, MSEB for audio, Julia 1 for decision-making.
3. Write a base prompt:Use the example models as a starting point.
4. Add constraints:Format, length, sources, performance metrics.
5. Execute and validate:Check output quality and tool-specific metrics.
6. Iterate:Based on feedback, modify the prompt and repeat the cycle.

6. Common mistakes and how to avoid them

• Lack of context:Add clear objectives, constraints, and output format.
• Prompt overloading:Break into multiple prompts or use chain-of-thought.
• Not testing on benchmarks:Use MSEB to validate audio encoders.
• Ignoring token budget:Set an explicit limit to avoid hidden costs.

Conclusion