When AI Agents Handle 80% of the Work, What Value Remains for Humans?

Compound Engineering Framework

The core of the Compound Engineering framework consists of four stages: Planning, Work, Review, and Compound. The first three resemble traditional software development, but the "Compound" stage adds systematic feedback of lessons learned back into the AI system, preventing repeated mistakes and enabling continuous improvement.

“If new insights are captured during the review, they are fed back to the system so the AI agent won’t repeat the same error next time—this is the most valuable part of Compound Engineering.”

AI Handles the Middle Execution

Kieran asserts that the Work stage can be largely delegated to AI agents, provided the Planning stage delivers a clear, detailed task breakdown and well‑defined boundaries. He emphasizes that LLMs can now run stably for hours or days, making the execution technically mature. However, handing over the planning to AI often yields vague solutions with hidden flaws.

“This isn’t blind trust; it’s a conclusion verified by practice: give the AI a precise plan and it will follow it rigorously.”

Human Core Interventions – The "AI Sandwich"

The "AI Sandwich" model likens humans to the two slices of bread and the AI agent to the filling. Humans must intervene at the beginning (Brainstorm and Ideate) to clarify the problem and generate diverse solution ideas, and at the end to refine quality, add aesthetic judgment, and make final decisions. The middle execution, testing, and iteration are left to the AI.

“Humans are the bread that determines the sandwich type; the AI is just the filling that adds texture.”

Why the Initial Phase Must Remain Human‑Led

Brainstorming (problem clarification) and Ideation (creative expansion) require nuanced judgment, cross‑domain insight, and the ability to switch cognitive frameworks—capabilities AI agents currently lack. AI can list options and fetch data, but selecting the most valuable direction is a human responsibility.

“Choosing which perspective to pursue is a core judgment that must be made by a human.”

Final Polishing Turns "Usable" into "Excellent"

After the AI completes tasks and passes automated tests, humans perform a last quality‑polish, akin to the Pomodoro technique’s remaining time. This step catches subtle usability issues and elevates a functional product to a compelling one, leveraging human aesthetic sense and attention to detail.

“Automated tests may show everything works, but a human’s hands‑on experience reveals minor frictions that, once fixed, turn a usable product into a great one.”

AGI Threshold and Future Outlook

Dan proposes that true AGI requires AI agents to run 24/7, be cost‑effective, and autonomously choose and prioritize new tasks without human prompts. Current agents can sustain long‑running single tasks but lack self‑directed planning and flexible cognition, indicating a long way to AGI.

“Before real AGI arrives, humans will remain the two ends of the sandwich, guiding direction and adding judgment that AI cannot replace.”

Implications for Software Engineers

Kieran notes that while engineers are still in demand, their role is shifting toward "product engineering"—overseeing the workflow, steering initial direction, and ensuring final quality, rather than writing every line of code. Mastering AI‑augmented processes becomes a core competency.

“Find your unique advantage and let the AI agent amplify it instead of replacing you.”