Agentic coding means using an AI agent to work inside your codebase. The agent can read files, plan changes, edit code, run tests, inspect output, and hand you a diff to review.

You can think of agentic coding tools as a junior engineer on your team. Give the agent the goal, constraints, and checks in plain English. It should read the code, ask questions when needed, break the work into steps, and make the change.

If you're new to agentic coding, pick one CLI agent or AI editor below, point it at a small, well-defined change in a real repository, and read the diff before you accept it. Completing a small real task will teach you more than another hour of reading.

Important note on safety: These AI agents run shell commands and change files on your machine. Use git or another version-control tool, review every terminal command before execution, review every diff before accepting, and keep permission prompts on. Do not start with YOLO mode. Once you understand how these agents behave, tune the workflow to fit your preferences.

Contents

Where Coding Agents Run

Models and Access Plans

Getting Better Results

Standards and Extensibility

Reference

CLI Coding Agents

CLI coding agents run in your terminal. They can read, edit, and run code in your repository.

  1. Claude Code - Anthropic's terminal-native coding agent for repository edits, shell workflows, tests, and pull-request preparation.

  2. OpenAI Codex CLI - OpenAI's open CLI coding agent for local and cloud-backed agentic development.

  3. OpenCode - Open-source terminal coding agent with a fast TUI, desktop app, and broad model-provider support.

  4. Pi - Minimal, hackable terminal agent you extend with packaged extensions, skills, and prompt templates.

  5. Crush - Charm's polished terminal agent, with multiple model providers, editor-grade code context, and MCP.

  6. Antigravity CLI - Google's CLI version of Antigravity agent, aimed at longer-running, multi-step work.

  7. GitHub Copilot CLI - GitHub's terminal-native Copilot agent, tied into Copilot subscriptions and GitHub workflows.

  8. Cursor CLI - Cursor's command-line agent for scripting and automation, including CI runs that respect your project rules.

More to explore:

AI Editors, IDEs, and Extensions

Editor-based agents put coding help where code navigation, visual diffs, inline completions, project rules, and human review already happen.

  1. Cursor - AI-first editor with an in-editor agent, whole-codebase context, and fast inline completion.

  2. GitHub Copilot - GitHub's AI layer across VS Code and GitHub.

  3. Devin Desktop - The former Windsurf, still usable as a single-agent editor (its multi-agent command center is under Agent Workspaces below).

  4. Google Antigravity IDE - Google's agent-first editor view, with in-editor agents, commands, and completions.

  5. Kiro - Agentic IDE and CLI built around spec-driven work, with hooks and project steering.

  6. Cline - Open-source agent that runs as an IDE extension, a CLI, or an embedded SDK.

More to explore:

Agent Workspaces and Orchestration

Agent workspaces and orchestration tools let you run agents in parallel, either in one codebase or across projects.

  1. OpenAI Codex app - Desktop command center for parallel Codex threads, worktrees, automations, and git review.

  2. Claude Code desktop app - A graphical Claude Code for running several sessions at once, with visual diffs and PR monitoring.

  3. GitHub Copilot app - GitHub's desktop app for running agents across issues and pull requests in parallel.

  4. Google Antigravity 2.0 - Platform for running agents across separate projects.

  5. Devin Desktop - Command center for running and tracking Devin agents, Kanban-style (built on the former Windsurf).

  6. Zed Agent Panel - Zed's built-in panel for supervising agent threads from one editor.

More to explore:

Hosted Models and Coding Subscriptions

To run coding agents with hosted models, you need access through a subscription or API. For individual use, a subscription is often better from cost perspective.

  1. ChatGPT and Codex plans - ChatGPT subscriptions cover Codex app and CLI access; the OpenAI API fits agents that accept API keys.

  2. Claude plans and Anthropic API - Common path for Claude Code, Claude subscriptions, and direct Anthropic API usage in custom agents.

  3. GitHub Copilot plans - Seat-based access to Copilot in editors, GitHub, Copilot CLI, and cloud coding-agent workflows.

  4. OpenCode Go - Low-cost OpenCode subscription for access to tested open-weight coding models through an API key usable by OpenCode or other agents.

  5. OpenCode Zen - Pay-as-you-go OpenCode model gateway for curated coding-agent models, including frontier and open-weight options.

  6. GLM Coding Plan - Z.ai subscription for GLM coding models usable from Claude Code, Cline, OpenCode, Kilo Code, and other compatible agents.

  7. MiniMax Token Plan - MiniMax subscription tiers for coding-agent and multimodal workloads, including MiniMax Code and API-platform usage.

  8. Cursor plans - Cursor's subscription, on a usage-based model, covering its editor, CLI, and cloud agents.

  9. OpenRouter - Unified model router for accessing many hosted models through one API surface with fallback and routing options.

  10. DeepSeek API - Low-cost direct API option for model-flexible agents that can use provider-specific or OpenAI-compatible keys.

More to explore:

  • CodingPlan.org - living comparison of coding subscriptions, model access, usage limits, and supported developer tools.

Run Models Locally

Instead of subscribing to a coding plan, you can run open-weight LLMs locally and use them with coding agents. This can help with privacy and cost. Expect local models to be slower, less capable than hosted models, and more dependent on your hardware.

  1. Ollama - Simple local model runner with CLI, desktop, server, and broad agent integrations.

  2. llama.cpp - Low-level, portable inference engine for GGUF models across CPUs, GPUs, and constrained hardware.

  3. vLLM - High-throughput inference and serving engine for shared GPUs, labs, and team-hosted model endpoints.

  4. LM Studio - Desktop and server workflow for discovering, running, and serving local models through chat and API surfaces.

More to explore:

Spec-Driven Workflows

If you ask an agent to start coding from a vague request, the result can drift from what you wanted. In spec-driven development (SDD), you write requirements, constraints, acceptance criteria, and edge cases first. Then the agent uses that spec to write code and tests.

  1. GitHub Spec Kit - Open toolkit for turning product scenarios into specs, plans, tasks, and implementation loops across coding agents.

  2. Agent OS - Standards and spec-writing system for bringing consistent project context to Claude Code, Cursor, and other agents.

  3. BMAD Method - AI-driven agile development framework that scales from bug fixes to larger product and architecture workflows.

  4. GSD - Lightweight context and spec system for Claude Code.

  5. Superpowers - Agent workflow made from small, reusable skills.

  6. OpenSpec - Spec-driven development framework for iterative AI-assisted work on existing repositories.

Agent Skills and Plugins

A skill is a reusable workflow you teach the agent once instead of re-explaining it in every prompt. For each skill, the agent reads a short description first and pulls in the full SKILL.md, plus any scripts or references, when the task calls for it.

  1. Agent Skills specification - The SKILL.md standard itself, and the place to start if you want to write your own.

  2. OpenAI Codex Skills - Official Codex docs for creating, installing, invoking, and distributing skills and plugins.

  3. Anthropic Skills - Reference examples for Claude-compatible skills, including technical, document, and creative workflows.

  4. Agent Skills by Addy Osmani - Skills for spec, plan, build, test, review, and release workflows.

More to explore:

UI Design With Coding Agents

LLMs often produce generic UI if you give them a feature request without design context. Give the agent concrete design context: Google Labs' DESIGN.md, reference screens, tokens, interaction rules, and a focused design-review skill such as Anthropic's frontend-design skill.

  1. DESIGN.md - File format for capturing visual identity, design-system rules, tokens, and reusable design decisions for agents.

  2. Anthropic frontend-design skill - Skill for building less generic frontend interfaces.

  3. Make Interfaces Feel Better - Agent Skill focused on spacing, alignment, motion, type rendering, and radius consistency.

  4. Awesome Claude Design - Ready-to-use design-system inspirations in DESIGN.md format for agent-built UI.

More to explore:

Standards and Interop

Each of these standards solves one integration problem, and it helps to know which.

  1. Model Context Protocol - Open protocol for connecting AI applications to external tools, data sources, and contextual systems.

  2. AGENTS.md - Simple repository-level instruction file for setup commands, tests, conventions, and agent behavior.

  3. Agent Skills - Open SKILL.md format for packaging reusable instructions, scripts, references, and assets with progressive disclosure.

  4. DESIGN.md - File format for giving agents persistent design-system context (see UI Design for putting it to use).

  5. Agent Client Protocol - Protocol for connecting editors and coding agents without locking each editor to one agent backend.

  6. Agent2Agent - Protocol for communication and interoperability between agentic applications.

MCP Servers

MCP is useful when an agent needs structured access to external tools or data. The official MCP docs explain the standard; the Codex MCP docs show concrete server configuration. For one-off work, a direct CLI or browser tool may be simpler.

  1. Context7 - MCP server and CLI that inject current library documentation and examples into coding-agent context.

  2. GitHub MCP Server - Official GitHub MCP server for repositories, issues, pull requests, and development workflows.

  3. Chrome DevTools MCP - Official Google MCP server that gives agents Chrome DevTools access for debugging and performance work.

  4. MCP reference servers - Reference and community servers for filesystem, git, memory, fetch, sequential thinking, and other core tasks.

More to explore:

  • awesome-mcp-servers - large directory of MCP servers across developer tools, data, browsers, cloud, and productivity.

Benchmarks

Use benchmarks to compare models and the agent harnesses that drive them.

  1. SWE-bench Pro - Software-engineering benchmark inspired by SWE-bench, aimed at realistic repository issues.

  2. Terminal-Bench 2.1 - Benchmark for evaluating agents in real terminal environments with validated tasks.

  3. DeepSWE - Benchmark for coding agents on long-running software tasks from active open-source repos.

  4. FrontierCode - Cognition benchmark focused on whether agent-produced code would meet production codebase and maintainer standards.

Learning and References

  1. AI Agents for Beginners - Microsoft course covering agent patterns, tools, multi-agent systems, and production practices.

  2. Hugging Face Agents Course - Free course on agent basics, tools, planning, multi-agent systems, and benchmarked assignments.

  3. MCP for Beginners - Microsoft curriculum for learning MCP fundamentals across .NET, Java, TypeScript, Python, and Rust.

  4. How to Build a Coding Agent - Workshop that builds a coding assistant from chatbot to file editing and code search.

  5. Martin Fowler CLI Coding Agent Walkthrough - Practical article on assembling a terminal coding agent around tools and local development standards.

  6. Build Your Own CLI Coding Agent from Scratch in Python - Self-guided workshop to build your own CLI coding agent like Claude Code from scratch in Python.

More to explore:

A Few Tips to Start

I'd recommend starting by asking the agent questions about a codebase you know before letting it edit anything. After that, give it small, clear tasks with a way to check the result, like fixing a failing test or a specific file to change. Then, add an AGENTS md in your codebase, with dev commands, style rules, and test instructions so you don't have to repeat yourself in every session. When a change is not what you expect and agent keeps going in the wrong direction, reset with git and try a better prompt instead of arguing with the agent. As you get comfortable, hand off bigger tasks and loosen the permission prompts.

Reply

Avatar

or to participate

Keep Reading

© 2026 Agentic Coding Weekly.
beehiivPowered by beehiiv