Methodologies and Frameworks

The industry is renegotiating what replaces story-driven delivery. The current frameworks try to redesign the whole system: manifestos on how humans and agents divide work, the spec-driven toolchain, SDKs for building agents and reference setups to copy from. The method for building an agent follows from what an agent is, a program wrapped around a model. Its working shape is a structured cycle that writes each lesson back, with a named person answering for what ships.

Start here

• The agent loop — A minimal model of an agent and the loop that turns a model into a working system.
• Spec-driven development — How specifications create stable intent and verifiable work for coding agents.
• The disciplined compounding loop — How structured cycles capture lessons and make each subsequent change cheaper.
• Human accountability — Where human judgment and responsibility remain essential in agentic workflows.

Related

• Agentic Engineering: the discipline these methods serve.
• Quality and Verification: the checks every method relies on.
• Memory and Context Engineering: where each cycle writes its lessons back.
• Moving away from Agile: McKinsey’s AI operating-model research: evidence on operating-model change beyond the tool.
• Product Management: how specifications and intent move closer to the product side.

Further reading

Two write-ups on how teams structure agent-assisted work in practice, both from The Refactoring (refactoring.fm) and worth reading before designing a workflow from scratch: How Engineering Teams use AI (opens in a new tab) (12 March 2025) and Vibe-coding workflows (opens in a new tab) (30 April 2025).

Resources

Methodologies & Manifestos

• 12-Factor Agents (opens in a new tab) - Twelve principles for building production LLM-powered software.
• AI Manifesto (opens in a new tab) - Five principles for responsible AI use in development, keeping human insight, accountability, and learning central rather than letting the model speak or think for you.
• AI-Augmented Software Development Manifesto (opens in a new tab) - Reworking of the Agile Manifesto for agents, valuing human taste, tight feedback loops, tested software, and adaptable intent over autopilot.
• ASDLC.io (opens in a new tab) - Knowledge base and framework for the Agentic Software Development Life Cycle, with agents as a logistics layer and humans governing intent; MIT, early stage.
• Agent-Native Architectures Guide (opens in a new tab) - Technical guide for apps where agents are first-class citizens, covering parity, granularity, composability, file-based interfaces, conflict models, and execution patterns.
• Applied GenAI (FINOS) (opens in a new tab) - FINOS initiative on GenAI in financial services with open-source tools, governance frameworks, the Open Financial LLM Leaderboard, and EMIR-specific RAG for regulatory compliance.
• Context Engineering (opens in a new tab) - First-principles handbook inspired by foundational context-engineering work and 3Blue1Brown, on moving past prompt engineering into context design, orchestration, and optimization.
• HumanLayer (opens in a new tab) - Framework that wraps AI coding agents with deterministic human-approval workflows for high-stakes operations.
• Glamorous Toolkit (opens in a new tab) - Moldable development environment: build a custom inspector per domain and compose views, so the effort goes into understanding a system rather than only writing it (talk (opens in a new tab)).
• Moldable Development (opens in a new tab) - Approach based on programming through contextual micro tools molded per problem, optimizing the largest cost in software, figuring the system out.
• PDCVR Framework (opens in a new tab) - Plan-Do-Check-Verify-Retrospect methodology for Claude Code with 25+ specialized subagents, TDD discipline, and structured workflows extending the traditional PDCA cycle.
• PROSE (opens in a new tab) - Framework of agentic primitives (skills, custom agents, instructions, hooks) for GitHub Copilot, paired with the Agentic SDLC Handbook.
• Ponytail (opens in a new tab) - “Lazy senior dev” ruleset that pushes coding agents toward minimal code before writing implementations; MIT, distributed as plugins for Claude Code, Codex, Cursor, and others.
• SuperClaude framework (opens in a new tab) - Configuration framework that adds specialized commands and personas on top of Claude Code.
• Swamp (opens in a new tab) - AI-native CLI for authoring reviewable, shareable operational workflows against typed models of external systems, with Claude Code skills bundled into the binary.
• Vibe Coding Playbook (opens in a new tab) - Methodology repo with prompting techniques, step-by-step implementation guides, and XML prompting patterns for AI-assisted development.
• o16g (Outcome Engineering) (opens in a new tab) - Sixteen-principle manifesto reframing agentic development around verified outcomes over code, managing to cost not capacity, and auditing agents as they drift.

Spec-Driven Development

• Agent OS (Builder Methods) (opens in a new tab) - Free system for defining and managing coding standards in AI development, discovering standards from your codebase and injecting them into spec shaping.
• Allium (opens in a new tab) - Specification language that captures behavioral specs alongside implementation to guide AI code generation and keep intent and behavior aligned.
• BMAD Method (opens in a new tab) - Breakthrough Method for Agile AI-Driven Development: 12+ AI personas plus structured workflows that walk a team through the SDLC from idea to deployment.
• CCPM (opens in a new tab) - Project-management skill system for agents that drives spec-driven development and parallel execution via GitHub Issues and Git worktrees.
• Codervisor Lean Spec (opens in a new tab) - Lightweight, flexible Spec-Driven Development (SDD) for AI-powered development.
• Conductor (Gemini CLI) (opens in a new tab) - Gemini CLI extension for spec-and-plan workflows with centralized context artifacts and version-control conventions for greenfield and brownfield projects.
• DESIGN.md (Google Labs) (opens in a new tab) - Format spec for describing visual identity to coding agents, combining YAML design tokens with prose rationale, plus a CLI for linting and WCAG contrast checks.
• Get Shit Done (GSD) (opens in a new tab) - Lightweight meta-prompting, context-engineering, and spec-driven-development setup for Claude Code by TACHES.
• GitHub Spec Kit (opens in a new tab) - GitHub’s toolkit for spec-driven development, scaffolding the workflow where written specs steer AI coding assistants.
• Intent (opens in a new tab) - Contract-based language that compiles a single .intent source to native Rust binaries, JavaScript, and WebAssembly, with pre/postconditions, Z3 SMT verification, and property-based test generation.
• Kiro (opens in a new tab) - Spec-driven IDE that bridges AI coding and engineering, handling long-running tasks across large codebases with an agent that learns how you work.
• OpenSpec (opens in a new tab) - Open-source spec-driven development tool by Fission AI for AI coding assistants.
• RequireKit (opens in a new tab) - AI requirements toolkit using EARS notation for structured specs, BDD/Gherkin generation, epic/feature hierarchy, and optional GuardKit integration.
• Rezonant (opens in a new tab) - Product workspace that carries product thinking from messy inspiration through to shaped, engineering-ready work for coding agents to build.
• Spec-Driven Role Prompts (opens in a new tab) - System prompts for a five-role spec-driven workflow with REQUIREMENTS.md to SPEC.md handoffs, traceable requirement IDs, and variants tuned for Claude Code, Codex, and earlier models.
• Tessl (opens in a new tab) - Spec-driven development tool with a Spec Registry of 10,000+ open-source library specs, giving coding agents structured versioned context.
• The Website Specification (opens in a new tab) - Platform-agnostic spec of 162 website best practices, the technical things a good website should do, written for humans and agents and queryable via an MCP server and llms.txt.

Agent SDKs & Frameworks

• AgentOS (opens in a new tab) - Pre-built FastAPI runtime that turns Agno agents, teams, and workflows into a scalable API server with orchestration, state, and observability (formerly github.com/buildermethods/agent-os (opens in a new tab)).
• Agno (opens in a new tab) - Agent framework paired with AgentOS for running multi-agent systems inside your own cloud.
• Amazon Bedrock AgentCore (opens in a new tab) - AWS framework-agnostic Python SDK for production agents with Runtime, Memory, Code Interpreter, Browser, Gateway, and Observability, plus the AgentCore CLI (opens in a new tab) for project scaffolding.
• Claude Agent SDK (opens in a new tab) - Anthropic’s Python (opens in a new tab) and TypeScript (opens in a new tab) SDK exposing the Claude Code engine as a library with agent loop, built-in tools, context management, and in-process MCP tools.
• CrewAI (opens in a new tab) - Standalone Python framework (no LangChain dependency) for role-playing autonomous agents, with Crews for collaboration and Flows for enterprise production.
• DSPy (opens in a new tab) - (stanfordnlp - GitHub (opens in a new tab)) Stanford NLP framework for programming, not prompting, LMs: write modular Python pipelines (classifiers, RAG, agents) and DSPy optimizes prompts and weights algorithmically.
• Eve (opens in a new tab) - Vercel’s filesystem-first TypeScript framework (Apache 2.0) for building durable agents, with instructions, tools, skills, channels, and schedules in conventional directories; public beta.
• Flue (opens in a new tab) - Astro team’s sandbox agent framework in TypeScript (Apache 2.0) for building durable autonomous agents, with secure sandboxes for code execution and file edits.
• Google Agent Development Kit (ADK) (opens in a new tab) - Google’s open-source code-first agent toolkit in Python (opens in a new tab), Go (opens in a new tab), and TypeScript/JS (opens in a new tab) with multi-agent composition and deployment to Cloud Run; model-agnostic but Gemini-tuned, with docs (opens in a new tab) and a built-in web dev UI (opens in a new tab) for debugging, tracing, and evals.
• IBM watsonx Orchestrate ADK (opens in a new tab) - Python SDK and orchestrate CLI for building, testing, and deploying agents on IBM watsonx Orchestrate, with a Docker-based local Developer Edition for offline iteration.
• LangGraph (opens in a new tab) - LangChain’s low-level orchestration framework for building stateful agents as graphs with durable execution, human-in-the-loop, and short/long-term memory; ships in Python and TypeScript (opens in a new tab).
• Mastra (opens in a new tab) - TypeScript framework from the Gatsby team for building AI applications and agents, covering model routing, workflows, human-in-the-loop, and production observability.
• Microsoft Agent Framework (opens in a new tab) - Microsoft’s SDK for .NET and Python that merges Semantic Kernel and AutoGen into one runtime with graph-based workflows; AutoGen now in maintenance mode.
• OpenAI Agents (opens in a new tab) - AGENTS.md, a simple open format for guiding coding agents, with examples and templates.
• OpenAI Agents SDK (opens in a new tab) - Lightweight, provider-agnostic framework for multi-agent workflows in Python (opens in a new tab) and TypeScript (opens in a new tab) with handoffs, guardrails, human-in-the-loop, and tracing across 100+ LLMs.
• Parlant (opens in a new tab) - Open-source Python framework (Apache 2.0) for customer-facing agents in regulated domains, using contextual matching to narrow LLM context to only the guidelines relevant per turn.
• Sandcastle (opens in a new tab) - TypeScript library for orchestrating sandboxed coding agents through a single sandcastle.run() call, with pluggable Docker, Podman, and Vercel sandbox providers.
• VoltAgent (opens in a new tab) - MIT-licensed TypeScript agent stack pairing the @voltagent/core framework with the VoltOps Console for observability, evals, prompts, and deployment.
• Zero (opens in a new tab) - Vercel Labs’ experimental systems programming language for building small native agent tools, with explicit effects, predictable memory, and structured compiler output; not yet stable.
• Plain-Java Claude Agent Framework (opens in a new tab) - Plain-Java agent framework that talks to Anthropic’s Claude with custom tool execution.

Examples & Templates

• ADK Sample Agents (opens in a new tab) - Ready-to-run agents built on Google’s Agent Development Kit in Python, TypeScript, Go, and Java, from simple bots to multi-agent workflows.
• AI That Works (opens in a new tab) - Weekly live-coding series from HumanLayer with per-episode code on production AI engineering, context engineering, and harness design.
• Agent Skills (opens in a new tab) - Production engineering skills for AI coding agents covering the full SDLC, installable as a Claude Code plugin and compatible with Cursor and Gemini CLI.
• Agentic Coding Starter Kit (opens in a new tab) - Template for AI web apps built to be driven by coding agents, shipping AGENTS.md, CLAUDE.md, a spec-based workflow, and optional agent skills. MIT-licensed.
• Augmented Coding Patterns (opens in a new tab) - Catalog of 43+ patterns, 9 anti-patterns, and 14 obstacles for collaborating with AI coding assistants.
• Auto Plan Mode (opens in a new tab) - Defensive system prompt that forces Plan Mode before destructive actions via –append-system-prompt.
• Beast Mode for VS Code (opens in a new tab) - Custom VS Code chat mode with autonomous workflow guidelines and research steps.
• Claude Code Workflows (opens in a new tab) - Curated Claude Code workflows and configs from an AI-native startup’s daily use.
• Claude MD Memory Workflow (opens in a new tab) - Claude Code memory-management workflow used in YouTube videos.
• Claude Managed Agents Demo (opens in a new tab) - Unblocked’s proof-of-concept showing autonomous agents on Anthropic’s Managed Agents API that pull organizational context through the Unblocked MCP server and open pull requests. MIT-licensed.
• Command examples (opens in a new tab) - Claude Code commands from an LLM agent that generates n8n automation workflows from a single prompt.
• Cursor Rules for Java (opens in a new tab) - Curated Skills and Agents for Java enterprise SDLC workflows, usable across AI agent harnesses.
• Claude Code Dotfiles Example (opens in a new tab) - Personal dotfiles with a Claude Code config to crib from.
• Claude Code Dotfiles Example (opens in a new tab) - Personal dotfiles with a Claude Code config to crib from.
• Living Architecture (opens in a new tab) - Monorepo example with Claude Code integration: automatic code review, custom commands, hooks, and structured agent workflows, surfacing architecture as living documentation.
• Medium Download example (opens in a new tab) - Example CLAUDE.md from a tool that downloads all blog posts from Medium.
• Memo (Gitpod) (opens in a new tab) - Notion-style note-taking app built entirely by AI agents in a public two-week sprint, streamed live to showcase an agent-driven SDLC with no human-written code.
• Microsoft Agent Framework Samples (opens in a new tab) - Hands-on Python and .NET tutorials for the Microsoft Agent Framework covering tools, RAG, workflows, evaluation, and multi-agent cases.
• Mike (opens in a new tab) - Open-source AI legal SaaS (AGPL-3.0) on Next.js + Express, Supabase auth/Postgres, S3-compatible storage, multi-model providers, DOC/DOCX-to-PDF via LibreOffice, and one-shot Supabase schema migrations.
• PlatformPlatform AGENTS.md (opens in a new tab) - Enterprise .NET monorepo demonstrating AI-guided development with DDD, CQRS, multi-tenant SaaS patterns, and 30+ automation workflows.
• Proxmox Manager (opens in a new tab) - Model-driven Proxmox VM lifecycle automation built on Swamp, composing TypeScript extension models with YAML workflows for stateful provisioning, PXE boot, and Discord-bot control.
• Requirements Analyst Agent (SHIFT to AI) (opens in a new tab) - Downloadable system-prompt files turning a coding agent into a requirements analyst that interviews you through nine discovery phases and emits a REQUIREMENTS.md before any code.
• Agent Skills Collection (opens in a new tab) - Agent Skills collection shared from a working .claude directory.
• TheLastRESTService (opens in a new tab) - AI-driven REST service that generates endpoints from request patterns and emits OpenAPI specs in real time.