An open-source knowledge repository, architecture framework, and specification suite for building, running, and constraining Agentic AI systems.
Traditional software executes fixed code. Chat AI generates passive text. Agentic AI combines reasoning models with execution environments to autonomously achieve goals through tool invocation, environment feedback, and iterative reflection.
Goal -> [Perceive -> Decide -> Act -> Observe -> Reflect] -> Verified Outcome
This repository collects core concepts, engineering disciplines, architecture blueprints, standard protocol specifications, reference code implementations, and production checklists for building enterprise agentic workflows.
- 📄 agentic-ai: What makes a system "agentic" — moving from passive prompts to goal-seeking loops.
- 📄 agentic-workflow: Dynamic decision loops vs linear DAG automation.
- 📄 agentification: Wrapping legacy APIs and CLIs into self-describing agent tools.
- 📄 agent-skills: On-demand instruction loading (
SKILL.md) and modular skill packages. - 📄 loop-engineering: Perception-action-reflection loops, exit criteria, cycle detection, and iteration budgets.
- 📄 graph-engineering: State graphs, dynamic branch routing, nodes, edges, and state transitions.
- 📄 context-engineering: Context window payload assembly, rule framing, dynamic slicing, and signal density.
- 📄 capability-engineering: Tool sandboxing, SafePath scoping, host API grants, and least capability principles.
- 📄 state-engineering: Working state management, checkpoints, transaction boundaries, and rollbacks.
- 📄 tool-calling: Structuring, type-checking, and executing LLM function calls safely.
- 📄 agentic-runtime: Execution environments engineered specifically for hosting agents (sandboxing, energy budgeting, SafePath).
- 📄 mcp: Model Context Protocol — standardizing client-agent tool interfaces.
- 📄 agent-memory: Context window management, sliding summaries, and Knowledge Item persistence.
- 📄 agentic-security: Indirect prompt injection defense, tool output poisoning, and runtime confinement.
- 📄 reflection: Self-critique, root cause log analysis, and iterative hypothesis revision.
- 📄 hallucination-mitigation: Source inspection grounding and empirical assertion gates.
- 📄 multi-agent: Multi-agent collaboration patterns (Supervisor, Specialist Workers, Delegation).
- 📄 prompt-routing: Routing tasks dynamically between fast models and heavy reasoning models.
- 📄 human-in-the-loop: Escalation boundaries and human verification checkpoints.
- 📄 agentic-governance: Audit trails, DLP policies, and fleet compliance frameworks.
- 📄 evaluation: Benchmark suites, step efficiency metrics, and assertion gates for AI agents.
- 📄 Autonomous Coding Assistant Blueprint: Architecture map of a pair-programming agent with planning, tool bridge, and test verification.
- 📄 Autonomous DevOps & SRE Agent Blueprint: Architecture map of a monitoring, diagnostic, and auto-rollback SRE agent.
- 📄 Multi-Agent Supervisor Engine Blueprint: Architecture map of a hierarchical supervisor agent managing specialized worker subagents.
- 📄 Enterprise MCP Gateway Server Blueprint: Architecture map of a central MCP discovery, ACL, and audit proxy server.
- 📄 Headless Browser Agent Kernel Blueprint: Architecture map for browser automation, accessibility DOM parsing, and video recording.
- 📄 Autonomous ETL & Data Pipeline Agent Blueprint: Architecture map for autonomous data extraction, cleaning, and normalized DB ingestion.
- 📄 Agent Skill Package Specification v1.0: Standard specification for
SKILL.mdfrontmatter, resolution order, and skill layout. - 📄 Agent Trajectory Transcript Format v1.0: JSON Lines (
transcript.jsonl) schema specification for auditing agent executions. - 📄 Tool Execution Contract v1.0: Standard contract for tool execution payloads, SafePath bounds, and async tasks.
- 📄 MCP Server Conformance Suite v1.0: Standard conformance suite, JSON-RPC validation, and error code criteria.
- 📄 Agent Memory & KI Format v1.0: Standard specification for persistent Knowledge Items and working memory frames.
- 📄 Capability Grant Protocol v1.0: Standard protocol for capability lifetime scopes, descriptors, and SafePath bounds.
- 💻 reference/loop.go: Production-grade reference implementation of an energy-metered, step-bounded Agentic Loop kernel in Go.
- 💻 reference/mcp_client.go: Clean Go client implementation for connecting agents to MCP servers via JSON-RPC.
- 💻 reference/context.go: Production Go reference implementation of context compaction, sliding tool output summaries, and prompt rules assembly.
- 💻 reference/safepath.go: Production Go reference implementation of SafePath workspace boundary validation and file confinement.
- 📄 Top 10 Agentic AI Anti-Patterns: 10 common anti-patterns in agent development and how to avoid them.
- 📄 Production Readiness Checklist: 15-point verification checklist before deploying agents into production.
- 📄 Indirect Prompt Injection Defense Guide: Practical strategy for isolating untrusted data inputs and preventing prompt overrides.
- 📄 Evaluating AI Agent Performance: Step-by-step guide to building benchmark datasets, step efficiency metrics, and regression testing.
Maintained as part of the WithAgenticAI ecosystem.
- 🌐 Website: withagentic.ai
- 🐙 GitHub: github.com/WithAgenticAI/AgenticAI
- 📄 License: MIT