Jeff Dean Explains Context Engineering

Google’s former Chief Scientist Jeff Dean says choosing the latest or most powerful AI model is becoming less important as developers focus more on how models are used within wider AI systems.

In a recent interview with Y Combinator’s Diana Hu, Dean discussed the growing importance of context engineering – the process of giving AI models the right information, tools and instructions to complete a task effectively.

Rather than relying on the model alone, developers can combine AI with retrieval systems, external tools, memory and other agents to create more capable systems.

AI Models Are Only One Part of the System

The rapid development of AI has often focused on bigger models, more training data and increased computing power. Dean suggests that the focus is now shifting towards what happens around the model.

He explained that an AI model is only one part of a larger system designed to solve a particular problem. Giving the model access to relevant information and the right tools can have a major impact on how well it performs.

This means developers do not necessarily need to keep switching to a newer model to improve results. In some cases, better results can come from improving the way the existing model receives information and interacts with other systems.

Retrieval and Tools Improve AI Performance

One important part of context engineering is retrieval. Instead of expecting an AI model to rely entirely on information contained within its training data, developers can provide it with relevant information for the specific task.

This can make the information much easier for the model to use because it is directly available within its working context.

For example, an AI system designed to answer questions about a company could retrieve the latest product information, internal documents or customer data before generating its response.

Tools can also allow AI systems to perform actions rather than simply generate text. Depending on the application, these could include searching databases, accessing APIs, analysing information or carrying out calculations.

The goal is to give the model everything it needs to handle the task rather than expecting the model itself to contain every answer.

Multi-Agent AI Systems Are Becoming More Important

Dean also highlighted the growing role of agent and multi-agent systems.

Rather than asking one AI model to complete an entire complex task, a system can break the problem into smaller steps. Different tools or AI agents can then handle individual parts of the process.

This requires careful orchestration. The system needs to determine which tools are appropriate, what information should be retrieved, which steps should happen first and whether the results are good enough.

In more advanced systems, an AI agent could try one approach, assess the outcome and then use another tool or strategy if the first attempt fails.

This type of coordination is likely to become increasingly important as AI moves beyond simple question-and-answer applications.

How Developers Can Improve Context Engineering

Dean’s advice for improving AI systems is largely based on experimentation.

Developers can start by using AI systems to tackle real problems and then examine where they fail. These failures can reveal what information, instructions or tools are missing.

Instead of immediately trying to change the underlying model, developers can often improve performance by creating clearer instructions or giving the model better ways to use available tools.

For example, if an AI repeatedly struggles with a particular type of task, developers could create a specific workflow or set of instructions to guide it through the process.

Over time, these improvements can create a system that becomes better at solving a particular class of problems.

Why Context Engineering Could Become More Important

The wider AI industry has spent years competing over model size, training data and benchmark performance. Dean’s comments suggest that attention is increasingly moving towards the systems built around those models.

A highly capable model can still produce poor results if it lacks the right information or does not know how to use the tools available to it.

On the other hand, a well-designed system can potentially get much more from an existing model by providing relevant context, reliable retrieval and carefully designed workflows.

This could also make AI development more accessible. Building a new foundation model requires enormous amounts of computing power, data and investment, while improving the context around an existing model can be achieved with far fewer resources.

Key Takeaways

Dean’s comments highlight a shift in how AI systems are being developed:

• The model is only one part of the system. The tools and information surrounding it can be just as important.
• Context matters. Giving an AI system relevant information for a specific task can improve its ability to produce useful results.
• Retrieval and tools are becoming essential. AI can be more effective when it can access information and perform actions when needed.
• Multi-agent systems are growing in importance. Complex tasks can be divided between different agents, tools and workflows.
• Failure can help improve AI systems. Developers can study where models struggle and use those insights to create better instructions and processes.

As AI development continues, the competitive advantage may increasingly come not from simply choosing the newest model, but from building a better system around it.