From Writing Code to Designing Agent Experience
Agent experience engineering is the discipline of designing, supervising, and improving AI agents so that human intent turns into reliable software outcomes without humans writing every line of code. Instead of focusing mainly on syntax and frameworks, engineers define goals, constraints, and feedback loops for systems that operate at machine speed. Netlify CTO Dana Lawson argues that writing code was never the bulk of engineering work and is now “increasingly the least strategic bit.” In her view, engineers become “shepherds of production,” responsible for understanding what goes in, what comes out, and how agents behave in between. This shift coincides with Lawson’s forecast of a billion new applications by 2029, created by both professional developers and new “builders” who express intent in conversational language instead of traditional programming languages.
AI Coding Agents Are Growing Up—and Need Oversight
AI coding agents are evolving from one-shot assistants into persistent collaborators that write tests, debug failures, generate documentation, and handle implementation workflows over time. Anthropic’s agentic coding forecast describes a world where agents compress implementation, testing, and iteration into much shorter loops, functioning like a junior team that never sleeps. Yet the company’s research shows that while developers use AI in roughly 60% of their work, they only delegate up to 20% of tasks completely. That gap reveals the new reality: machines do more of the labor, but humans still carry the responsibility. Agent experience engineering is about closing that gap safely by setting architecture, acceptance criteria, and review standards so that coding agents speed progress without turning into an accountability trap.
Why Managing Machines Beats Chasing Speed
The next competitive edge in software will come from AI agent management, not from buying the latest AI tools or chasing speed alone. Evidence from GitHub and METR shows that AI can make developers faster in some settings and slower in others, depending on workflow design and verification costs. Google Cloud’s DORA research frames AI-assisted development as a systems problem: fast local coding can create downstream chaos in testing, security, and support if oversight is weak. Companies that treat AI coding agents like unsupervised automation risk rewarding volume over quality and hiding rework. The winners will treat agents as accelerators inside carefully designed systems, where human attention focuses on what matters most while agents handle routine checks and raise boundary cases.
New Software Developer Skills for an Agentic Era
As AI coding agents take on more implementation work, software developer skills are shifting toward system thinking, product sense, and risk judgment. Lawson describes agent experience as combining developer experience and user experience, extended to include AI agents as a first-class audience. Engineers must decide what to build and, more importantly, what not to build in a world where human bandwidth is no longer the main bottleneck. That includes shaping clear intent, designing event-driven agent workflows, and ensuring that outputs align with business, security, and ethical constraints. Outcome-focused engineering practices become the anchor: engineers define success upfront, then use agents to explore options quickly while guarding against a flood of low-quality or obsolete software that fails users and organizations.
The Rise of the AI Supervisor Across the Organization
Agent experience engineering will not stay inside the engineering team. Anthropic expects non-technical groups in sales, marketing, legal, and operations to use agentic tools to build workflows with little direct help from developers. Stack Overflow’s survey found that 84% of respondents use or plan to use AI tools in their development process, even though more developers distrust AI accuracy than trust it. That tension is shaping a new role: the AI supervisor. Whatever their title, these people define goals, constrain agents, inspect work, test outcomes, and decide when to stop. For engineering leaders, that means building guardrails: escalation rules, human approval gates for sensitive actions, automated tests that agents cannot bypass, and audit trails that show who authorized what.
