Deep Agents and the Evolution of AI Workflows: A Discussion with LangChain and Box

Defining AI Agents and Their Evolution

• 💡 AI agents are technically defined as an LLM running in a loop, calling tools until a task is complete.
• 📈 The concept has evolved from simple LLM calls and chatbots to more complex workflow-like systems and now "deep agents."
• ⚠️ Early agents often "spiraled out" due to models not being advanced enough, but current models are significantly better.

From Workflows to Deep Agents

• 🔄 Agents differ from traditional workflows because the agent itself determines when to proceed to the next step, often using LLMs for decision-making and conditional logic.
• ⚙️ Frameworks like LangGraph represent agentic workflows as graphs with cycles, allowing for more control and complex interactions.
• 🚀 Deep Agents represent a new paradigm, building on tool-calling loops with added features like built-in planning tools and access to file systems.

Key Components of Deep Agents

• 📝 Complex prompts are crucial, as they encapsulate much of the system's complexity, often including detailed instructions and tool descriptions.
• 💾 File systems are vital for managing state, offering flexibility in data formats (e.g., JSON) and allowing LLMs to structure their working memory.
• 🎯 Deep agents excel in complex, read-heavy tasks like research and coding, often producing "first drafts" that require human review.

User Experience and Time Considerations

• ⏱️ The user experience (UX) for interacting with agents is an underexplored area, especially regarding user patience and response times.
• ⚡ Users expect synchronous interactions to be very fast for "in-flow" augmentation, while asynchronous tasks allow for longer processing times to replace more substantial work.
• 🧠 There's an "uncanny valley" of time where tasks taking a few minutes are too long for synchronous flow but too short to be considered substantial asynchronous work.

Evaluating Agent Performance

• ✅ Simple agent tasks can be evaluated using input/output eval sets to check for specific tool triggers or expected outcomes.
• 🧐 For complex, long-running agent trajectories, human review becomes essential, as a simple "yes/no" evaluation is insufficient.
• 💡 Future evaluation could involve agents building their own eval sets or leveraging user feedback and memory to learn from mistakes and improve over time.

Future of AI Workflows

• 🌱 The field is in a state of continuous change, but core skills like prompting, tool descriptions, and retrieval remain foundational.
• 🏗️ New frameworks and capabilities build upon existing ones, creating a layered stack of AI technologies.
• 🚀 Users and companies are encouraged to experiment and learn now, as the insights gained will remain valuable even as the technology evolves.