The best AI solution is mostly deterministic software, with LLM judgment layered in
When deciding how a redesigned workflow should operate once intelligence is built in, you need to answer four questions: Where does deterministic software live? Where does the agent act? Where does the human approve? Where does the record get updated? The best solution for most companies is a strong combination of deterministic software — probably the majority of the system — plus the judgment that API calls to LLMs can provide, and then a human-in-the-loop layer on top. The agent handles the first half — intake, validation, drafting — a human approves, and then the latter half of the steps execute automatically. This architecture is what a good FDE should push for in every implementation.
Go deeper with Greg Isenberg
Ask anything about this, and Greg Isenberg answers from everything they've shared.
More to learn
See everything in Learn
MembersHow Hyperagent builds a skill: it researches style first, then distills it
When you ask Hyperagent to create a skill — for example, one that posts in the style of a specific creator like Greg Isenberg — it doesn't just write a prompt that mimics the person. Instead, it draws up a plan: first research the creator's
MembersSkills are the most important primitive in frontier AI agents
Skills are described as the single most important concept in the frontier-agent world. The underlying models are already generally intelligent — think of it like hiring an Albert Einstein who is brilliant in the abstract but doesn't know re
The $999 AI Assessment Playbook: Sell, Run, and Upsell a Productized AI Audit
Product: The $999 AI Assessment Playbook: Sell, Run, and Upsell a Productized AI Audit Format: Digital download: the-999-ai-assessment-playbook-sell-run-and-upsell-a-product.pdf Price: $79.00 USD Description: Learn to land your first pay