# Focus: What I'm working on now

> "Focus is not about saying yes to one thing, but saying no to a thousand other good ideas."

Currently, my work is centered around three pillars of the AI Agent ecosystem: exploring the theoretical boundaries of Agent capabilities, defining the methodology for building robust Agent workflows (EDF), and delivering high-leverage outcomes that validate these ideas.

---

## 1. Agent Capabilities & Ecosystem (The Future)

*Exploring the architectural boundaries of proactive intelligence.*

I am researching how to move beyond simple chat-based interactions towards truly proactive agents that understand intent, maintain long-term memory, and navigate complex context windows.

### Context Engineering

- **Passive Context Management**: System Prompts, context, CoT, agent.md, soul.md, toolset.md, Meta-layer context management
- **Active Context Management**: MCP, Skill, CLI, API - Agents proactively sensing their environment
- **Memory Systems**: Persistent memory layers based on ACT-R cognitive architecture
- **Proactive Harness**: Framework for proactive action based on environmental cues rather than reactive prompts

### EDF Convention

- **Guardrails**: Using Core DAGs and Finite State Machines (FSM) to ensure safety
- **Interface Guidelines**: Recommended interaction patterns for Agent-driven UIs
- **Service Contracts**: Standardizing how Agents interact with external APIs

---

## 2. Optimizing Workflows with EDF (The Present)

*Methodology for the next generation of product development.*

The **Executable Design File (EDF)** framework is my core methodology. The central thesis is: **Use a robust DAG/FSM as the Guardrail, and let the Agent serve as the Engine.** This approach bridges the gap between deterministic business logic and the creative flexibility of LLMs.

- **Efficiency**: Drastically accelerating the design-to-production pipeline by reducing manual coding for standardized patterns
- **Reliability**: Ensuring agents operate within bounded autonomy, making them safe for production environments
- **Scalability**: Decoupling the intent logic from the implementation, allowing for rapid iteration of both

### What EDF 2.0 means

In traditional workflows, requirements, design files, and code are isolated artifacts. EDF 2.0 puts them into one Git repo using Markdown, JSON / YAML state machines, API contracts, design metadata, code, and tests, making product context readable and operable by Agents.

It is not another PRD template, and it is not just a Design-to-Code tool. It is closer to a repo-native product convention: product judgment, business logic, interface constraints, implementation, and validation stay aligned inside one shared context.

### EDF as a single source of truth

In this system, SSOT is not an abstract slogan. It is a set of readable repo assets:

- **Why**: product judgment, user problems, business logic
- **What**: state flows, business rules, decision trees
- **Contract / Data**: APIs, fields, entity constraints
- **Interface**: design surfaces, component semantics, state expression
- **Src / Test**: implementation, validation, and regression protection

### The two flywheels

EDF 2.0 is meant to accelerate two flywheels:

- **Time to Insight**: forming judgment faster from user, market, experiment, and technical signals
- **Time to Value**: turning judgment into runnable, testable, iterative product value faster

```text
Time to Insight Flywheel
Signal / Why → Judgment / What
User, market, experiment, and technical signals → product judgment

             ↓ drives

Time to Value Flywheel
Interface → Prototype / Src → Feedback / Test
Interface expression → runnable implementation → feedback and verification

             ↺ feeds back

New feedback and experiment results return to Time to Insight
```

### Three working principles

1. **Any-Point Entry**  
   Work does not always begin from a PRD. It can begin from a screenshot, design draft, code prototype, GTM copy, user feedback, or market signal.

2. **Reverse Extraction**  
   Once a page or prototype runs, an Agent can extract state, contracts, and design constraints back from code and interface behavior.

3. **Guardrails**  
   Tests, contract checks, cross-layer diffs, and human review keep implementation from drifting away from the original product judgment.

---

## 3. Practical Outcomes (Launched / Active)

*Tangible proof-of-concepts and tools built using the EDF methodology.*

### 📂 Presenter Agent

**A new form of presentation service.**

Traditional slides are linear and static. Presenter Agent transforms presentations into **navigable information spaces**. Using multi-level zoom (Overview → Detail → Deep dive), it allows speakers to navigate contents dynamically based on audience feedback.

- *Status*: Prototype Active (EDF validation case)

### 📂 Xiaohongshu Long-image Agent

**Real-world validation of the EDF framework.**

A tool designed to automate the creation of high-quality long-form images for social media. It serves as a benchmark for the EDF methodology: proving that a "Core DAG + Agent Engine" pattern can replace traditional manual development workflows for content-heavy applications.

- *Status*: Requirement Definition & Early implementation

---

> [!TIP]
> Interested in technical details? Check out the [Playground](https://nickmu.com/en/playground) for deep dives into Cognitive Architectures and EDF specifications.