From Agent Prompter to Agent Manager

The biggest hurdle to AI adoption is not how smart the models are; it is the human tendency to over-engineer how we orchestrate them. Alessio Fanelli shows that the real advantage comes from treating AI as a managed process. This means using stable state machines like Linear instead of relying on fragile, human-in-the-loop chat interfaces. By shifting from prompting to managing, you stop babysitting individual tasks and start building systems that scale. This is the only way to move from a cool demo to a useful business tool. The advantage goes to those who treat AI as a persistent, state-aware worker rather than a reactive chatbot.

The Architecture of Autonomy

Conventional wisdom in AI development pushes for complex, custom orchestration layers. Fanelli’s experience suggests the opposite: the most effective systems are often the simplest, relying on existing project management tools.

The State Machine Advantage

Most teams fail at agentic workflows because they treat the AI as a conversational partner. When you put an agent in a chat loop, you lose the ability to track progress, handle rework, or audit costs. Fanelli uses Linear as a state machine. The agent does not chat; it monitors a board. When a task moves to To Do, the agent executes. When it hits Human Review, the human intervenes.

I think people over-engineer at first what these things can be and ultimately the power of LLMs, especially these newer models is you could just give them a spec for how they will work and they will lock to that spec when executing whatever you hand it.

-- Alessio Fanelli

This creates a persistent, auditable record. If a task fails, you are not scrolling through a chat history; you are looking at a structured workflow with a clear original spec and rework checklist.

The Hidden Cost of Magic Instructions

There is a common trap of building skills files or system prompts that grow indefinitely. Fanelli warns that models tend to add rather than remove instructions. Over time, your CLAUDE.md or skills.md becomes a bloated, contradictory mess that confuses the agent.

The fix is counter-intuitive: purge your instructions. If you find yourself adding a line to tell an agent not to do something, you are likely masking a deeper issue in your tooling or the agent environment. A clean, minimal spec is more durable than a comprehensive one.

Scaling Through Physical Intersection

The most interesting consequence of AI is its ability to make small, unscalable businesses efficient. Fanelli’s Pokemon card scouting agent, which autonomously browses eBay, extracts certificate data, and flags underpriced inventory, is a masterclass in using AI to bridge the gap between digital data and physical assets.

There are a lot of businesses that are based on kind of highly heterogeneous data that have been impossible to scale with software because before you have kind of something as malleable as an LLM that can go through these things. It is really hard to use even like text or image classification for these things.

-- Alessio Fanelli

This is not just about automation; it is about capturing value that was previously lost to human bandwidth constraints. When you automate the searching and pricing of inventory, you are not just saving time; you are expanding the total addressable market of your business.

Key Action Items

• Audit Your Orchestration (Immediate): If you are using a chat-first approach for multi-step tasks, migrate to a state-based system like Linear or Jira. If the agent cannot move a ticket from To Do to Done without you holding its hand, it is not an agent; it is a glorified autocomplete.
• Purge Your System Prompts (Next 30 days): Review your skills.md or system instructions. If you have lines that start with don't do X, delete them. If the model is failing, simplify the goal rather than adding more guardrail text.
• Track Token Spend by Task (Ongoing): You cannot optimize what you do not measure. Assign a ledger to your agentic tasks. If a 10-million-token task costs more than the value of the output, stop trying to fix the prompt and start fixing the tooling.
• Implement Senses for Agents (Next 3-6 months): If your agents struggle with web-based tasks, do not just add more instructions. Integrate tools like Playwright or visual scrapers like Glimpse to give the agent better senses. Visual feedback often solves problems that text-based prompting cannot.
• Identify Heterogeneous Bottlenecks (12-18 months): Look for areas in your business or personal life where you manually reconcile data across different formats, such as physical cards versus eBay listings or inventory lists versus physical freezers. These are the high-leverage areas where AI creates the most value.