This skill provides a structured methodology for building applications and systems driven by AI agents. Unlike traditional development, agent projects require an approach focused on managing uncertainty, validating model capabilities, and building robust and idempotent pipelines.
Mental Model: Don't automate by intuition; automate by evidence. The success of an agent project depends on finding the "Task-Model Fit" and building an infrastructure that allows iteration on prompts and data without breaking the execution flow.
🚩 Fragility and Autonomy
- Fragility: High. Model performance drift and ambiguity in input data can invalidate an architecture in days.
- Freedom: Medium in the choice of tools, but Low in the discipline of validation and structuring of pipelines.
🚀 When to Activate
- When starting a new project that involves autonomous agents or LLM workflows.
- To evaluate whether a task is suitable to be automated with current technology.
- When designing the architecture of a batch processing or massive data analysis system.
- To rescue agent projects that have failed due to lack of structure or excessive costs.
🧠 Agentic Project Lifecycle
1. Discovery & Task-Model Fit
Is the task suitable for the current model?
- Manual Prototyping: Perform the task manually in the chat. Can the model do it consistently?
- Complexity Assessment: Does it require multi-step reasoning, tool access, or just synthesis?
- Boundary Definition: Define what the agent will NOT do to avoid "scope creep" and catastrophic failures.
2. Pipeline Architecture (The Canonical Flow)
Divide the task into deterministic stages whenever possible:
- Acquire: Obtaining raw data (Web, DB, Files).
- Prepare: Cleaning and formatting for the LLM context.
- Process: The only stage that involves calls to the LLM (Gemini 3).
- Parse: Extraction of structured data from the response.
- Render: Generation of the final output (UI, Report, Code).
3. Filesystem as State Machine
Use the filesystem to maintain the task state. Each stage should read from one folder and write to another, allowing restarts and atomic debugging.
🛠️ Implementation in Gemini Elite Core
Pattern: Architectural Reduction
As models (e.g., Gemini 3 Ultra) become more capable, simplify the architecture.
- Before: You needed 5 agents and a supervisor for a technical summary.
- Now: A single prompt with good "Context Engineering" can achieve the same with lower latency and cost.
Pattern: Idempotency and Caching
Ensure that re-running the pipeline does not waste tokens if the inputs have not changed.
# Ejemplo de flujo de trabajo idempotente
bun run pipeline.ts --stage process --batch-id 2026-03-02
🔗 Integrations
evaluation: Validation of each stage of the pipeline.filesystem-context: Implementation of state storage.context-optimization: Maximization of the signal in the 'Process' stage.
📚 Internal References
references/case-studies.md: Analysis of real projects (Market analysis, Massive refactoring).references/pipeline-patterns.md: Technical patterns of parsing, error handling, and cost estimation.