Skip to content

Latest commit

 

History

106 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Learning Agentic AI

A long-running learning and portfolio repository focused on applied AI systems, agentic AI, backend AI engineering, and controlled tool-use harnesses.

Each module is an isolated, runnable project managed with uv. The repository has evolved from early framework explorations into a structured portfolio track - Agent Factory - that demonstrates production-oriented harness design for AI tool execution.

Project thesis:

Agents are useful only when the application harness controls identity, validation, policy, approval, execution, and audit.


Repository Structure

learn-agentic-ai/
├── 00-uv/                          # uv package manager fundamentals
├── 01-lite-llm/uv-proj/             # LiteLLM unified LLM API gateway
├── 02-crewai/00-hello-crewai/      # CrewAI multi-agent framework
├── 03-agent-factory/               # Main portfolio / system-building track
│   ├── 00-about-and-thesis/
│   ├── 01-ai-prompting-2026/
│   ├── 02-how-to-think-ai-era/
│   ├── 03-agentic-coding-crash-course/
│   ├── 04-build-ai-agents/
│   └── xx-projects/                # Numbered artifact sequence
│       ├── 00-identity-aware-agent-harness
│       ├── 01-llm-proposed-skill-runner
│       ├── 02-approval-gated-github-tool-harness
│       ├── 03-durable-side-effect-ledger
│       ├── 04-approval-gated-real-github-comment-adapter
│       ├── 05-real-mode-smoke-evidence-release-gate
│       └── 06-operator-approval-workbench

03-agent-factory is the main portfolio and system-building track. It contains coursework, thesis notes, and the xx-projects/ artifact sequence - a series of progressively more capable controlled-agent harnesses.


Agent Factory - Artifact Sequence

The Agent Factory track builds a series of numbered artifacts, each adding a new layer of controlled AI tool execution.

# Artifact Path Status Core Claim
00 Identity-Aware Stateful Agent Harness 00-identity-aware-agent-harness Complete / preserved Server-derived identity, role/scope policy, stateful task lifecycle, approval gates, audit trail, LangGraph checkpoint/resume
01 LLM-Proposed, Harness-Controlled Skill Runner 01-llm-proposed-skill-runner Complete / tagged artifact-2.2 Model-shaped skill proposals, proposal validation, policy and approval lifecycle, Skill Runner API, validated scalar arguments, safe rejection of unsafe/control-plane/malformed args
02 Approval-Gated GitHub Tool Harness 02-approval-gated-github-tool-harness Complete as local/demo fake-client artifact One approval-gated GitHub issue-comment skill path, validated scalar arguments, trusted repository policy, explicit approval, side-effect idempotency through an in-memory ledger, fake-client execution, audit evidence, adversarial safety tests
03 Durable Side-Effect Ledger and Approval Binding 03-durable-side-effect-ledger Complete as local/demo durable fake-client safety artifact SQLite-backed side-effect records, durable approval binding, restart/replay duplicate suppression, durable audit evidence, fake-client execution
04 Approval-Gated Real GitHub Comment Adapter 04-approval-gated-real-github-comment-adapter Complete as local/demo real-comment adapter Fake-client default plus explicitly configured allowlisted real GitHub issue-comment path with durable approval binding, token boundary, remote marker lookup/reconciliation, and audit evidence
05 Real-Mode Smoke Evidence and Release Gate 05-real-mode-smoke-evidence-release-gate Complete / published / tagged evidence artifact Release-gate evidence context for Artifact 04, with redacted manual smoke evidence, offline replay/no-duplicate proof, and zero-network negative proof
06 Operator Approval Console / Workbench 06-operator-approval-workbench Current local/demo workbench artifact (A6.5) Operator-facing approval inbox, explicit operator approve/reject routes, local static workbench, status/audit/ledger visibility, and demo packaging

Key Design Boundaries

  • Artifact 00: Request bodies cannot claim identity, role, or scopes. Identity is server-derived.
  • Artifact 01 / 2.2: The proposer proposes. The harness validates, authorizes, approval-gates, executes, and audits. Dry-run only; scalar args only; no real GitHub writes; no live LLM HTTP mode.
  • Artifact 02: Local/demo fake-client only. No real GitHub API calls. Not production-ready.
  • Artifact 03: Durable local/demo fake-client safety artifact with SQLite side-effect records, approval bindings, restart/replay duplicate suppression, and durable audit events.
  • Artifact 04: Local/demo real GitHub comment adapter. Fake client remains default; real mode is explicit, allowlisted, token-boundary-controlled, and narrow to one issue-comment operation.
  • Artifact 05: Evidence and release-gate context only. It does not own runtime code or broaden GitHub operations.
  • Artifact 06: Current local/demo operator workbench. Fake/default execution only for the demo; no live GitHub, token, .env, Next.js, package-managed frontend, deployment, OAuth/OIDC, or production console claim.

Current Leading Artifact

Artifact 06 is the current leading artifact. It turns the approval-gated harness lineage into a local/demo operator workbench: a proposed action appears in an approval inbox, the operator reviews risk/scopes/context/execution mode, approves or rejects through server-controlled A6 routes, and then inspects status, decision history, local/demo audit timeline, side-effect/ledger visibility, and execution-result evidence.

Artifact 06 derives its runtime baseline from Artifact 04. Artifact 05 is referenced as release-gate evidence context only.

Artifact 06 remains local/demo and fake/default by default. It does not claim production readiness, production authentication, deployment, arbitrary GitHub automation, or live GitHub execution for the demo.


Recommended Review Path

  1. Start with this root README.
  2. Open 03-agent-factory/README.md.
  3. Open 03-agent-factory/xx-projects/README.md.
  4. Review Artifact 6 README.
  5. Review Artifact 6 project status.
  6. Review Artifact 6 operator workbench demo.
  7. Review A6.5 demo evidence.

Current Limitations

This repository contains learning and portfolio artifacts, not a deployed production service.

Current Artifact 06 limitations:

  • Local/demo operator workbench only
  • Fake/default execution for the demo
  • No live GitHub execution required or performed by the demo
  • No GitHub token or .env required for the demo
  • No OAuth/OIDC production identity provider
  • No deployment or production authentication
  • No Next.js frontend, package.json, or node_modules
  • No MCP integration
  • No arbitrary GitHub automation claim
  • Not production-ready

Early Learning Modules

The early modules remain as reference material from the initial learning phase.

00-uv - uv Package Manager

Project Description
00_hello_world Minimal uv project - pyproject.toml, virtual environments, script entry points
01_blog_flow_uv Blog generation pipeline built with CrewAI Flows, managed via uv

01-lite-llm - LiteLLM

Project Description
uv-proj Calls Google Gemini 2.0 using LiteLLM's completion() API

02-crewai - CrewAI

Project Description
00_hello_crewai First multi-agent crew - agents with roles, goals, tasks

Technologies Used

Technology Purpose
uv Fast Python package and project manager
FastAPI API framework for agent harness endpoints
LangGraph Stateful graph execution with checkpoint/resume
Pydantic Strict domain model validation
pytest Test framework for all artifacts
Ruff Python linter and formatter
CrewAI Multi-agent orchestration (early modules)
LiteLLM Unified, provider-agnostic LLM API (early modules)
Python >= 3.12 Language runtime

Prerequisites

  • Python >= 3.12
  • uv installed:
    curl -LsSf https://astral.sh/uv/install.sh | sh

Author

Harry5174 - harisjaved010@gmail.com

About

Hands-on agentic AI projects exploring tool execution, approval gates, durable state, audit trails, and reliable side effects.

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages