This framework is divided into two parts. The first part is the strategic filter for identifying a viable business opportunity. The second part provides the core principles for designing, building, and scaling the product itself.
Part 1: The Strategic Filter (Finding the Right Problem)
This section helps you decide if a business idea is worth pursuing in the first place.
1. Solve a "High-Friction" Workflow, Not a Single Task
What it means: Don't just automate one simple action. Target a complex process that requires a human to act as the "glue" between different tools, tabs, or modes of thinking. These "swivel-chair" tasks, like sales prospecting or research synthesis, are where agents provide exponential value.
Why it matters: The value of an agent is proportional to the complexity and frustration of the workflow it replaces. Solving a multi-step, high-friction problem creates a "must-have" solution, not just a "nice-to-have" feature.
2. The "Human-in-the-Loop" is a Feature, Not a Bug
What it means: Aim to make a human 10x more efficient, not to replace them. The agent's primary job is to produce a high-quality "first draft" that a human can then quickly approve, edit, or reject.
Why it matters: This approach builds trust, ensures quality control, and makes the product immediately usable, even with the current limitations of AI. It turns the human from a manual labourer into a skilled editor, which is a powerful and desirable value proposition.
3. Your Moat is Your Workflow & Your Data, Not the LLM
What it means: Anyone can access a powerful LLM. Your defensible business comes from two things:
- (A) Your deep, expert understanding of a specific business workflow, encoded into the agent's logic
- (B) Your ability to leverage a customer's proprietary data via RAG to give the agent unique knowledge
Why it matters: A business built only on a clever prompt is not defensible. A business built on a proprietary workflow and a unique data advantage creates a powerful moat that gets deeper the more customers you have.
4. Start Vertical, Not Horizontal
What it means: Don't build a generic "research agent." Build a "research agent for patent lawyers" or a "due diligence assistant for M&A analysts." A specific, vertical customer has a specific, painful problem and a willingness to pay for a tailored solution.
Why it matters: A vertical strategy focuses your product development, simplifies your marketing, and allows you to build a solution that is far superior to any generic tool for that specific niche. It's the classic "inch-wide, mile-deep" approach.
Part 2: The Product & Engineering Principles (Building the Right Solution)
This section guides you on how to build a product that users will trust, adopt, and love, while ensuring it's technically and financially viable.
5. Nail the 'Job to be Done' and the User Interface
What it means: Deeply understand the user's existing process and integrate your agent into that flow as seamlessly as possible. The best agentic products don't force users into a new chat window; they meet them where they are, a button in Salesforce, a Slack command, an integration in their code editor.
Recommended by LinkedIn
Why it matters: The user interface is the product. A powerful agent that is hard to use will fail. Success depends on making the agent an effortless and natural extension of the user's current workflow, not a disruption to it.
6. Design for "Glass Box," Not "Black Box"
What it means: Show your work. The user interface should allow the user to easily inspect the agent's reasoning. This could mean showing the sources it used, the steps it took, or the key data points that led to its conclusion.
Why it matters: Trust is the single biggest barrier to AI adoption. A "glass box" design demystifies the AI, makes its output verifiable, and builds the long-term user confidence necessary for the product to become an indispensable part of their toolkit.
7. Prioritise Controllability and Reliability over Raw Intelligence
What it means: An agent that is 80% accurate but 100% predictable is more valuable than one that is 95% accurate but occasionally produces wildly incorrect or poorly formatted output. Use techniques like strong output parsing (e.g., forcing JSON output), rule-based guardrails, and fixed templates to constrain the agent's behaviour.
Why it matters: Business processes require reliability. You must be able to guarantee that the agent's output can be safely passed to the next step in a workflow, whether that's a human or another system. Consistency is often more important than brilliance.
8. Manage the Unit Economics from Day One
What it means: Every run of your agent costs money in API calls. You must design your agent's workflow to be as efficient as possible. This involves:
- Using a "router" to select the right model for the job (e.g., a cheap, fast model for simple tasks; a powerful, expensive model for complex reasoning)
- Aggressively caching results where possible
Why it matters: A business is not viable if the cost to serve a user exceeds the revenue they generate. Managing your LLM-related costs is a core engineering challenge that directly impacts your pricing, profitability, and ability to scale.
9. Establish a Robust Evaluation Framework
What it means: Before you change a prompt or update a model, you need a way to measure if you're making things better or worse. This means creating a "test suite" of representative problems and using automated metrics (and human review) to score the agent's performance on them.
Why it matters: This is how you move from "prompt engineering" to "AI engineering." A strong evaluation framework allows you to iterate quickly and confidently, provides quality assurance, and is essential for proving the reliability and safety of your system to enterprise customers.
10. Build a Virtuous Feedback Loop
What it means: Every action a user takes on the agent's output (editing, approving, rejecting) is an invaluable piece of feedback. Your product must be designed to capture this data and create a pipeline to feed it back into the system.
Why it matters: This is how your moat deepens. This feedback is the fuel for improving your prompts, refining your RAG documents, and potentially fine-tuning your own models over time. It creates a product that gets smarter with every user interaction, compounding your competitive advantage.
11. Abstract the Model Layer: Your Logic is the Asset
What it means: The underlying LLM is a commodity that will improve and change. Architect your application so that your core business logic is separate from the specific model you are calling.
Why it matters: This future-proofs your business. It allows you to easily swap in better, cheaper, or more specialised models as they are released, ensuring you always have access to the best technology without being locked into a single provider. It reinforces that your true asset is the workflow you've built, not the model you use.