Simon's Decision Framework: A Simple Structure for Complex Decisions

We make thousands of decisions every day, but when the stakes are high, we need a better approach than just “winging it.” Herbert Simon, a Nobel Prize-winning economist and cognitive scientist, gave us a deceptively simple framework that works whether you’re managing a crisis, building an AI system, or just trying to make better choices in your daily life.

His framework has three phases: Intelligence, Design, and Choice. Think of it as “Figure out what’s happening,” “Come up with options,” and “Pick one and do it.” Simple enough, but there’s real power in this structure.

The Three Phases Explained

Intelligence: Understanding the Situation

This isn’t about being smart-it’s about gathering intelligence, like a spy collecting information. You’re trying to understand what’s actually happening and whether you need to make a decision at all.

Say you’re running a team and notice productivity has dropped. The Intelligence phase means you gather data: Is it seasonal? Are people overwhelmed? Is there a tool that’s not working? Did a key person leave? You’re not jumping to solutions yet-you’re just figuring out what the real problem is.

In technical terms, this is your situation assessment. You’re scanning the environment, collecting relevant information, and identifying the need for action. Sometimes the “decision” is realizing you don’t need to decide anything yet.

Design: Creating Options

Once you understand the situation, the Design phase is about generating possible responses. This is where creativity meets analysis. You want to come up with several viable options, not just the first one that pops into your head.

Back to that productivity problem: Maybe you could redistribute workload, hire contractors, improve tools, adjust deadlines, or reorganize the team structure. The goal isn’t to pick the best option yet-it’s to make sure you have good options to choose from.

This phase often gets rushed, but it’s crucial. The quality of your final decision depends heavily on the quality of the options you generate here.

Choice: Deciding and Acting

The Choice phase is where you select your best option and implement it. But it’s not just about picking-it’s about committing to action and monitoring results.

You evaluate your options against your criteria (cost, time, risk, likelihood of success), make the call, and then actually do something about it. Crucially, you also set up ways to track whether your decision is working and adjust if needed.

Why This Framework Works

Simon’s framework addresses a fundamental truth about human decision-making: we’re not perfectly rational. We don’t have complete information, unlimited time, or infinite processing power. Simon called this “bounded rationality”: we make decisions within constraints.

The framework helps by:

Preventing Rush Decisions: The structure forces you to actually understand the problem before solving it. How many times have you solved the wrong problem really well?

Expanding Options: By dedicating a whole phase to generating alternatives, you avoid the trap of only considering obvious solutions.

Balancing Speed and Quality: The framework provides structure without being rigid. You can move through the phases quickly when needed, but you don’t skip essential steps.

Managing Complexity: Breaking decision-making into phases makes complex situations more manageable. Instead of trying to do everything at once, you focus on one phase at a time.

Simon’s Framework in LLM Agent Systems

Here’s where things get interesting for anyone building modern AI agent systems. Simon’s framework maps perfectly onto how we design LLM-based agents, and understanding this connection can make your agent systems more effective and reliable.

Intelligence Phase = Information Gathering

In LLM agent systems, the Intelligence phase is all about gathering the information needed to understand what’s really going on:

• Retrieval Systems: Searching through knowledge bases, documentation, or previous conversations to find relevant context
• Web Search: Using search engines to find current information that wasn’t in the training data
• User Queries: Asking clarifying questions when the initial request is ambiguous or incomplete
• API Calls: Fetching real-time data from external systems like databases, weather services, or stock prices
• Context Analysis: Understanding the user’s intent, constraints, and what success looks like

Think of this as your research agent-the one that goes out and gathers all the pieces of the puzzle before anyone tries to solve it.

Design Phase = Solution Generation

The Design phase is where multiple agents can work in parallel to generate different approaches:

• Specialist Agents: Different agents with expertise in different domains proposing solutions from their perspectives
• Creative Generation: Brainstorming multiple approaches without immediately judging them
• Research-Based Options: Using the gathered intelligence to craft solutions grounded in facts
• Constraint Checking: Ensuring proposed solutions are actually feasible given the user’s limitations
• Alternative Strategies: Generating backup plans or different approaches to the same goal

Instead of one agent trying to do everything, you might have a technical agent, a creative agent, and a practical agent each proposing their best solution.

Choice Phase = Decision and Response

The Choice phase brings everything together with a coordinator agent:

• Option Evaluation: Systematically comparing the proposed solutions against the user’s criteria
• Risk Assessment: Understanding the tradeoffs and potential downsides of each approach
• Final Decision: Selecting the best option or combining elements from multiple proposals
• Response Generation: Crafting a clear, actionable response for the user
• Follow-up Planning: Setting up ways to check if the solution worked and adjust if needed

This coordinator agent acts like a project manager, weighing all the input and making the final call.

Bounded Rationality in LLM Agents

Just like humans, LLM agent systems operate under real constraints. They have limited context windows, work with imperfect or outdated information, face API rate limits, and often need to respond quickly. Acknowledging these limitations leads to better agent design.

Instead of trying to gather perfect information or generate perfect solutions, you design agent systems that find good-enough answers quickly and reliably. This might mean setting search limits, using cached results when fresh data isn’t critical, or having agents satisfice rather than optimize.

Building Better Systems (Human or LLM Agent)

Whether you’re improving your own decision-making or designing LLM agent systems, Simon’s framework suggests some practical approaches:

For Human Decision-Making:

• Take time to really understand problems before solving them
• Generate multiple options before choosing
• Set up feedback loops to learn from your decisions
• Accept that you’ll rarely have perfect information

For LLM Agent Systems:

• Separate information gathering, solution generation, and decision-making into distinct agent roles
• Build in multiple solution generators, not just single-shot responses
• Include feedback mechanisms to learn from user satisfaction
• Design for real-world constraints like token limits and API costs from the start

A Framework, Not a Formula

Simon’s framework isn’t a rigid process-it’s a way of thinking about decisions. Sometimes you’ll cycle through the phases multiple times. Sometimes you’ll spend most of your time in one phase. The key is having a structure that helps you avoid common decision-making traps while remaining flexible enough to handle the messy reality of complex problems.

Whether you’re building the next generation of AI systems or just trying to make better choices in your work and life, this simple three-phase structure provides a solid foundation. Intelligence, Design, Choice-understand, create options, decide and act. Simple concepts, powerful results.