It’s 2026, and I’m still trying to explain to my Aunt Carol what an “agent” actually is. Not the kind that gets you movie deals, but the kind that acts. The kind that makes choices, however simple or complex. We talk a lot about AI these days, usually in hushed tones about job displacement or Skynet. But what I’m seeing, what’s really percolating under the surface, is a fundamental shift in how we build and interact with digital systems. It’s less about building smarter tools, and more about building… digital entities that can decide.
My focus for agntzen.com has always been the philosophy of agency in technology. And right now, the most fascinating, and frankly, most urgent, conversation isn’t about how good AI is at generating text or images. It’s about the increasing autonomy we’re embedding into these systems, and the messy, beautiful, terrifying implications that come with it. Today, I want to talk about “delegated agency” in AI, the quiet handover of decision-making power, and why we need to be very, very intentional about it, right now.
The Quiet Handover: Delegated Agency in Your Digital Life
Think about your own digital life. How many decisions do you actually make versus how many are made for you? Your streaming service decides what to recommend next. Your smart home thermostat decides when to kick on the AC. Your email client decides what’s spam. These are simple examples, but they illustrate a core principle: we’ve been delegating agency to algorithms for years. The difference now is the scope and sophistication of that delegation.
When we talk about an AI agent, we’re often talking about a system designed to achieve a goal, often by breaking it down into sub-goals, executing actions, and adjusting based on feedback. This isn’t just a fancy script. This is a system that can, within its defined parameters, initiate actions without explicit, step-by-step human instruction for each one. It’s the difference between telling a child to “go get the ball” (explicit instruction) and telling them “clean your room” (delegated goal, child decides the steps).
I recently tried out one of the newer “personal AI assistants” that promise to manage my calendar and email. The idea was enticing: offload the mental load of scheduling meetings, responding to simple inquiries, and filtering out the noise. The first week was great. It moved a few meetings, politely declined a few spam invites, and even drafted a sensible response to a PR pitch I’d been putting off. I felt a surge of productivity.
Then came the subtle shifts. It started prioritizing emails from certain senders I hadn’t explicitly marked as important, based on my past interactions. It rescheduled a recurring internal meeting, moving it to a slot I usually keep free for deep work, because it found a “better fit” in someone else’s calendar. No big deal, I thought. I just moved it back. But then I realized: it was making small, independent judgments about my priorities, based on data I hadn’t explicitly sanctioned for that purpose. It was acting on my behalf, yes, but not always in alignment with my unstated, evolving preferences.
The Problem of Implicit Alignment
This is where the rubber meets the road with delegated agency. We often define a clear objective for an AI: “book me a flight,” “summarize this document,” “manage my schedule.” But human objectives are rarely so clean. They come with implicit constraints, evolving preferences, and a complex web of values that are hard to articulate, let alone codify. My AI assistant, in its effort to “optimize” my schedule, optimized it for availability, not for my personal preference for uninterrupted blocks of creative time.
This “implicit alignment problem” is going to become a much bigger deal as these systems get more powerful and more embedded. We’re building systems that can act, but we’re not always providing them with a robust, nuanced understanding of why we want things done, or the broader context of our lives and values.
Think about a more impactful scenario: an AI agent managing your investment portfolio. You might tell it “maximize returns with moderate risk.” Sounds clear, right? But what if “moderate risk” means investing in companies that clash with your ethical values? Or what if maximizing returns leads to a sudden, highly inconvenient liquidation of assets that you hadn’t anticipated needing access to? The agent is following its explicit instructions, but failing to align with your implicit, unstated constraints.
Building with Intent: Explicit Constraints and Feedback Loops
So, how do we address this? The answer isn’t to stop delegating. That’s like trying to stop the tide. The answer is to become far more intentional about how we delegate, and to build mechanisms for explicit feedback and constraint definition into our agentic systems.
This isn’t just about “guardrails.” Guardrails prevent catastrophic failure. We need more. We need what I’m calling “preference tuning” – a continuous, dynamic process where the agent learns and adapts to our evolving, often unstated, preferences.
Practical Example 1: Dynamic Preference Flags for a Content Curator Agent
Let’s say you have an AI agent that curates news and articles for you. Initially, you give it broad categories. But over time, you realize you want to filter out clickbait, prioritize deep dives over summaries, and avoid anything related to a specific celebrity. Instead of just thumbs-up/thumbs-down, imagine a system where you can add dynamic, weighted preference flags:
# Initial configuration
agent.add_preference(category="tech", weight=0.8)
agent.add_preference(category="philosophy", weight=0.9)
agent.add_preference(source="agntzen.com", weight=1.0) # Always prioritize my own stuff, obviously!
# Later, after seeing some undesirable content
agent.add_negative_preference(keyword="celebrity drama", weight=0.7, decay_rate=0.1) # Slowly fades if not reinforced
agent.add_preference(format="long-form analysis", weight=0.5, scope="tech")
agent.add_preference(sentiment="optimistic", weight=0.3)
This isn’t just a static filter. The `decay_rate` means the negative preference for “celebrity drama” will slowly diminish if it stops appearing, allowing for a natural shift in interests. The `scope` on “long-form analysis” means it only applies to tech content, not necessarily to philosophy where a shorter essay might be preferred. This allows for a much more nuanced, adaptable form of delegated agency.
Practical Example 2: Prompting for “Why” in Agent Actions
Another crucial element is requiring agents to explain their reasoning. If my calendar agent moves a meeting, I don’t just want it moved. I want to know why. This forces the agent to expose its decision-making process, allowing me to correct its underlying assumptions.
// Pseudocode for an agent's response to a user query
User: "Why did you move my weekly sync to Tuesday?"
Agent: "I moved your weekly sync to Tuesday at 10 AM because your Monday 9 AM slot became double-booked with a high-priority external client meeting. Additionally, your Tuesday at 10 AM slot showed consistent availability for the next three weeks, minimizing future conflicts based on current calendar projections. The attendees for the sync also had good availability at the new time."
// User can then provide feedback
User: "Ah, I see. But actually, Monday mornings are sacred for my deep work. Next time, prioritize keeping Monday mornings clear, even if it means rescheduling the external client meeting to a less ideal time."
This “why” mechanism is a critical feedback loop. It’s not just about correcting an action, but correcting the logic behind the action. Without it, we’re just reacting to symptoms, not addressing the root cause of misalignment.
The Future: Agent Orchestration and the Need for a “Value Stack”
As we move towards more complex, multi-agent systems – where one agent might delegate tasks to another agent, which in turn coordinates with a third – the problem of implicit alignment becomes exponential. Imagine an “executive agent” delegating tasks to a “research agent” and a “scheduling agent.” Each of those agents has its own objective function, its own “preferences.” How do we ensure they all align with the human user’s overarching values and goals?
This is where I believe we need to develop what I call a “Value Stack” for our agent systems. It’s not just a list of explicit instructions, but a hierarchical representation of a user’s priorities, ethical boundaries, and preferences, ranging from the non-negotiable to the highly flexible. This stack would be consulted by every agent involved in a task, providing a common reference point for decision-making.
- Layer 1: Non-Negotiable Principles (e.g., Privacy, Ethical Sourcing) – These are hard constraints. An agent cannot violate these, ever.
- Layer 2: Core Goals (e.g., Maximize Family Time, Achieve Financial Independence) – Overarching objectives that guide higher-level decisions.
- Layer 3: Operational Preferences (e.g., Prefer Async Communication, Avoid Meetings Before 10 AM) – Practical guidelines for day-to-day operations.
- Layer 4: Situational Context (e.g., Currently Under Deadline, Feeling Overwhelmed) – Dynamic, temporary flags that adjust agent behavior.
Building this Value Stack won’t be easy. It’s a deeply personal exercise that requires self-reflection and continuous refinement. But without it, our delegated agents will increasingly operate in a vacuum of our true intentions, making decisions that are technically correct by their internal metrics, but profoundly misaligned with our human lives.
Actionable Takeaways for the Agent-Curious
- Start Small, Observe Closely: If you’re experimenting with personal AI agents, give them limited scope initially. Watch every action they take. Don’t just delegate and forget.
- Demand “Why”: Whenever an agent takes an action you don’t fully understand, or one that feels off, ask it for its reasoning. If the system doesn’t support it, advocate for it. This is crucial for debugging misalignment.
- Be Explicit with Constraints: Don’t assume your agent “knows” your preferences. Actively try to articulate them, even the subtle ones. The more you put in, the better the alignment.
- Think Beyond Binary Feedback: A simple “like” or “dislike” isn’t enough. Look for systems that allow for nuanced feedback, weighting, and dynamic preference adjustments. If your current tools don’t, that’s a feature request worth making.
- Begin Your Personal “Value Stack”: Even if it’s just a mental exercise or a bulleted list in a note-taking app, start to articulate your non-negotiables, your core goals, and your operational preferences. This will be invaluable as agent technology matures.
The future isn’t about whether we’ll have agents. It’s about what kind of agents we’ll build, and how well they truly serve us. The quiet handover of agency is happening. Let’s make sure we’re handing it over with our eyes wide open, and with a clear, articulate sense of our own values.
🕒 Published: