Building Reliable Agentic Systems: Eno Reyes

Summary

This talk from Factory explores the development of 'droids'—autonomous systems for various software engineering tasks. It delves into the characteristics of agentic systems, including planning, decision-making, and environmental grounding. The speaker discusses strategies like pseudo common filters, subtask decomposition, and model predictive control to enhance autonomous systems' performance. The talk also touches on decision-making techniques, tool use, and the importance of human guidance in refining these advanced AI systems.

Takeaways

• 🤖 The mission of Factory is to bring autonomy to software engineering through the development of 'droids', which are autonomous systems applied to various stages of the software development lifecycle.
• 🧠 Droids have separate cognitive architectures tailored to specific tasks, such as code review, documentation, and end-to-end coding tasks, including refactoring and feature work.
• 📋 Key characteristics of an agentic system include planning, decision-making, and environmental grounding, which are essential for the system to interact effectively with its surroundings.
• 🔍 The pseudo common filter is used to maintain consistent reasoning throughout the execution of a plan, especially useful in long processes like code migration.
• 🔑 Subtask decomposition can lead to positive downstream effects by providing more fine-grained control over the planning process, though it may introduce more decisions for the system to make.
• 🔄 Model predictive control allows for adaptive replanning based on real-time feedback during the execution of a plan, which is crucial in rapidly changing environments.
• 📝 Explicit plan criteria can significantly improve the quality of agentic systems by defining successful structures or initial states for plans, despite the trade-off of reduced generalizability.
• 🤝 Consensus mechanisms in decision-making, such as prompt ensembles and cluster sampling, can increase accuracy but may introduce longer inference times.
• 💡 Explicit and analogical reasoning helps reduce decision-making complexity by outlining the system's reasoning process, making it easier for the system to make choices.
• 🔧 Environmental grounding involves tool use and building AI computer interfaces to streamline actions and workflows, which is particularly effective in domains with many development tools.
• 🔮 Bounded exploration is important for agents to gather context about the problem space, but finding the right balance of exploration time requires evaluation and can significantly impact the success of a plan.

Q & A

What is the mission of Factory as described in the script?

Factory's mission is to bring autonomy to software engineering by building autonomous systems known as 'droids' that are applied to different stages of the software development life cycle, including code review, documentation, testing, and end-to-end coding tasks.

What are the three characteristics that define an 'agentic system' according to the script?

The three characteristics that define an 'agentic system' are planning, decision-making, and environmental grounding.

What is the pseudo common filter and how does it relate to agentic systems?

The pseudo common filter is a concept inspired by control systems and robotics. It involves pinging intermediate reasoning as an agentic system works through a plan, allowing individual decisions in the plan to converge towards consistent reasoning.

How does subtask decomposition impact the planning process in agentic systems?

Subtask decomposition is a method where larger tasks are broken down into smaller, more manageable subtasks. Increasing the resolution or fidelity of these subtasks can provide finer control and clarity in defining the action space, but it also introduces more decisions for the system to make, potentially complicating the decision-making process.

What is model predictive control and how does it apply to agentic systems?

Model predictive control is a method of evaluating outcomes of subtasks and the current state, enabling adaptive replanning based on real-time feedback during the execution of a plan. This is particularly useful in environments that change rapidly or where there are other active human engagements.

Why are explicit plan criteria important in building agentic systems?

Explicit plan criteria are important because they define successful structures or initial states for plans, leading to strong downstream effects. They help in error reduction and maintaining successful trajectories for the agentic system.

What is the role of consensus mechanisms in decision-making for agentic systems?

Consensus mechanisms play a role in decision-making by ensuring self-consistency and accuracy in the system's decisions. Techniques like prompt ensembles and cluster sampling can be used to select optimal samples from multiple inferences, leading to higher quality decisions.

How does analogical reasoning assist in simplifying decision-making for agentic systems?

Analogous reasoning, also known as Chain of Thought or checklist prompting, involves the system explicitly outlining its reasoning or decision-making criteria. This reduces the complexity of the decision-making process by breaking it down into a series of simpler, more manageable steps.

What is the significance of fine-tuning in decision-making for agentic systems?

Fine-tuning is significant because it allows for the customization of a model to make specific decisions more effectively. It can improve the quality of decisions, especially for those that are out of distribution, by training the model with high-quality data relevant to the specific decision-making tasks.

How does simulation contribute to decision-making in agentic systems?

Simulation contributes to decision-making by allowing the system to sample multiple decision paths and simulate their outcomes, both with real and imagined execution paths. This helps in evaluating the potential results of different decisions and choosing the optimal path.

Why is environmental grounding important for agentic systems?

Environmental grounding is important because it allows the agent to interact with and reason about its external environment, which is critical for implementing agentic systems effectively. It involves the ability to read and write to these environments, which is a key part of the agent's functionality.