Basics of Agentic AI
Core concepts of autonomous agentic systems: tools, short-term memory, procedural skills, company context, and specialized agent profiles.
Agentic AI transforms Large Language Models from passive chat assistants into proactive, autonomous operators capable of executing complex workflows, interacting with computers, and completing multi-step business deliverables.
1. Core Architectural Pillars
The Valstorm agent runtime operates on 5 foundational concepts:
1. Tools
Tools provide agents with real-world execution capabilities:
- Interacting with the operating system and terminal shell.
- Querying and mutating database records (
valstorm_sql_query,valstorm_record_cud). - Reading, writing, and searching Cloud Knowledge Vaults (
valstorm_vfs_write_file,valstorm_vfs_search). - Triggering multi-modal vision inspection (
analyze_image).
2. Memory
Memory provides state persistence across interactions:
- Short-Term Session Memory: Preserves multi-turn conversation context and tool outputs within an active session.
- Declarative Long-Term Memory: Persistent user facts, role preferences, and workspace constraints stored via
memory_manage.
3. Skills
Skills are structured, step-by-step Standard Operating Procedures (SOPs) written in markdown:
- Instruct agents on best practices for complex tasks (e.g. Test-Driven Development, PR Workflows, Literature Reviews).
- Transform ad-hoc LLM reasoning into repeatable, deterministic operations.
4. Context Vaults
The Virtual File Service (VFS) acts as the company's long-term collective brain:
- Contains architecture specs, policies, pricing sheets, and historical logs.
- Hybrid search (dense vector embeddings + metadata) retrieves verified excerpts to ground agent outputs and eliminate hallucinations.
5. Profiles & Multi-Agent Delegation
Profiles combine specific tools, system instructions, memories, and skills tailored to distinct business functions:
- A Frontend Developer profile carries Playwright, React, and AST tools.
- A Chief of Staff profile carries CRM mutations, VFS authoring, and executive planning skills.
- The Orchestrator Pattern enables a primary agent to decompose complex goals and delegate micro-tasks to specialized sub-agents in isolated context windows.
2. Prompting & Delegation Best Practices
- State Clear Goals and Constraints: Explicitly define the desired output format, target file paths, and business rules.
- Provide Context Locations: Direct the agent to relevant VFS vaults or database collections.
- Use Procedural Skills: When executing standardized tasks, instruct the agent to inspect the relevant skill using
skill_view(name='<slug>'). - Leverage Human-in-the-Loop Gates: For high-risk, destructive, or irreversible actions, ensure the agent requests confirmation.