What are the current best practices for coding agents?
Current best practices for coding agents center on structured context management, skill-based specialization, and iterative agentic loops. Corroborated practices include using CLAUDE.md files to document project conventions 178, building context with 7 components to boost productivity 8× 182, employing agentic loops (plan → act → verify) for self-correction and autonomous shipping 262,518, delegating roles across models for cost/performance 125, and visually managing agents 266. The tooling landscape is anchored by Claude Code, with Hermes 321,372, iFixAi 105, and Freellmapi 110 as notable supporting tools. Many emerging tools and practices lack broad corroboration and should be tested cautiously.[10]
- Add a CLAUDE.md file documenting project conventions, rules, and context to prevent over-engineering and hallucinated APIs 178.
- Build context with 7 components—memory, instructions, examples, files, previous actions, tool results, state—to increase Claude Code productivity 8× 182.
- Employ agentic loops: plan → act → verify → repeat for self-correction; Anthropic writes >40% of code this way 262 and Claude Code’s loop/goal features enable overnight autonomous shipping 518.
- Start with predefined workflows and add autonomy only when the task demands it 227.
- Use Claude Code to automate video editing—jump cuts, captions, chapter titles, etc. 220.
- Install the Find Skills skill to discover and apply the best skills for any goal 183.
- Use OpenAI’s codex-plugin-cc to let Claude Code delegate tasks to Codex in the same terminal 141.
- Delegate roles across models: Fable for design, Opus for heavy reasoning, Sonnet for execution to optimize cost and performance 125.
- Use loops instead of single prompts to automate work 112.
- Point Claude Code at an Obsidian vault to turn notes into a searchable wiki 111.
- Use Freellmapi to aggregate free tiers of 16 LLM providers behind one local API 110.
- Run iFixAi to red-team agent configurations with 45 tests and independent grading 105.
- Rely on Claude Fable 5 to generate complex interactive visualizations from code without external assets 103.
- Use the ai-website-cloner template to reverse-engineer live websites into clean code 428.
- Build agent businesses by first doing the work manually, then automating the learned process 381.
- Run multiple Claude Code sessions in TMux with agent teams and Git worktree isolation for adversarial code reviews and parallel feature testing 376.
- Optimize Hermes agent costs by model choice, background tasks, and context usage 372.
- Let Hermes Agent learn repetitive workflows from a single demonstration and execute them autonomously 321.
- Set up a pixel office: turn each Claude Code session into a visual character with speech bubbles, file tracking, and permission visualization 266.
- Have Claude Code read and modify an open-source trading bot’s code to create a profitable automated strategy 446.
- Use TouchDesigner, MediaPipe, and Claude Code to build real-time hand-controlled 3D applications 180.
- Copy-paste from Anthropic’s official prompt library for task-specific prompts 176.
- Use Claude Code with Sonnet 5 to produce a professional website in 18 minutes 132.
- Use Claude Code to automate cross-market spread detection across 50 markets 173.
- Use Claude Code to generate a complete ad in 2 minutes at $0.80 cost (116, emerging).
- Use Claude Code to build a workflow that records iPhone camera motion via ARKit and passes it to Seedance for AI video generation (143, emerging).
- Use Claude Code to identify YouTube niches with little competition by analyzing search data (419, emerging).
- Use Claude Code to autonomously run a full content strategy loop for Instagram growth (442, emerging).
- Use Claude Code to generate a complete game that runs on a custom $50 handheld device (619, emerging).
- Use Claude Code to generate a functional VST plugin from a single prompt in under 2 hours (566, emerging).
- Use Claude Code to automate Reddit scraping and analysis for customer pain points and ad copy (720, emerging).[32]
Core tool: Claude Code 111,112,125. Specialized agents: Hermes learns workflows from one demo 321 and can be cost-optimized 372; ECC bundles 60 agents and 231 skills (177, emerging). Security: iFixAi red-teams agents 105. Cost reduction: Freellmapi unifies 16 free LLM tiers 110; converting code to PNG images can cut token costs by ~80% (230, emerging). Multi-agent orchestration: codex-plugin-cc 141, Orca (137, emerging), TMux + Git worktree 376, ruflo (519, emerging), Cmux (578, emerging), Commonly 2.0 (617, emerging). Visual management: pixel office setup 266, CNVS (322, emerging). Browser automation: Browser Use CLI 3.0 (217, emerging). Website cloning: ai-website-cloner 428. Other emerging tools: Dotcode 130, Morph 124, Agent-Reach 114,443, Lev8 117, Ghost MCP 483, Google ADK CLI (172, emerging), Paperclip (174, emerging), OpenAlice (122, emerging), TradingAgents-CN (123, emerging), VideoAgent (574, emerging), OpenMontage (657, emerging). Many tools are early and lack extensive corroboration.[31]
Many practices remain emerging with only single-source corroboration and should be treated as unverified. These include: using handoff skills to compress session context 226, providing a design system before generating screens 225, building an AI agent in under 10 minutes 224, using Claude and AI agents to generate a project brief and booking chatbot 222, Browser Use CLI 3.0 reducing token usage by 6x 217, learning AI agent design patterns like Reflection and Self-Correction 216, using 6 specific plugins for video creation 215, converting long text to images to cut token costs by 70% 185, building an AI meeting workflow with Cronofy 184, combining Remotion and Claude Code for $1.33 animated shorts 181, ECC configuration system 177, solar-powered mesh node with Ollama 175, Paperclip control plane 174, Google ADK CLI 172, Claude Code building ARKit-to-Seedance workflow 143, organizing commands by workflow categories 142, RULER reward engineering 140, expanding capabilities with skills like inbox triage 139, Claude Cowork with Gmail 138, Orca parallel agent tool 137, single skill file auditing Google Ads 136, Fable one-shot CLI agent 135, multiple AI agents in one conversation 133, Viktor.com free credits 131, Dotcode secure environment 130, Anthropic running 99% of engineers with agent swarms 129, Hermes Agent Dashboard custom plugins 128, $400 rig of GTX 1080s replacing subscriptions 127, AiToEarn content automation , Morph free requests , TradingAgents-CN multi-agent simulation , OpenAlice trading agent , Claude Code power user workflows , shot-scraper video recording , solo founder raising $2M with Claude Code , surveyor combining Claude Code with LiDAR , Lev8 outperforming Exa and Codex , OpenAI Codex agentic loop handling 40% of coding , Agent-Reach single command access , site with 1000+ landing page templates , future of AI coding as designing loops , repos for RAG inputs and voice cloning , Claude Fable 5 generating interactive visualizations , cross-verification with two agents , AI tools like Lev8 for prospecting , AgentReach MCP server , Claude Code content strategy loop , loops improving success rate , studying 10 GitHub repos for AI engineering , equipping Codex with specialized skills , Fable 5 generating browser games , AI companies open-sourcing internal patterns , production prompting as workflow design , MCP connecting Claude Code with external tools , GitHub Spec Kit , vibe-coding 3D games , GitHub repo of free AI APIs , structured GitHub repo skill for context , Claude Code identifying YouTube niches , Tokenmaxxing workflow , Claude Code as starting point , modern computer vision workflow , Codex subagent harnesses with verification loops , structured slash commands , best AI agent businesses starting manual , Agent One workflow for D2C ads , combining web search and memory in Hermes , integrating design system like Astryx , running multiple Claude Code sessions in TMux , AMD Ryzen AI Max+ for local models , assigning roles via SOUL.md and MEMORY.md , Hermes cost optimization , rebuilding /plan skill , Claude building language learning app , Google OKF standard , Claude design skills forcing direction , CNVS visual orchestration , Hermes learning workflows , Codex OAuth , Claude monitoring BTC with simulations , shadowing jobs to build agents , Plan-Build-Judge loop , Claude Fable for GTA 6 scripts , Claude Code plugin for financial modeling , Claude Fable for final code review , automating 50% of busywork with Claude Desktop , providing about-me and memory files , combining Obsidian and self-learning skill , file hierarchy for Claude Code , using capable model to generate skills , pixel office setup , teenager building multiplayer game , visual workspace with agent characters , stack of agents and MCPs , agentic loop used by Anthropic , creating project with persistent folder , Codex app improvements , Codex multi-agent workflows , all 6 elements of structured prompt , switching between chat, cowork, code modes , converting code to PNG images , Claude Code with 30+ AI video models , OpenMontage agent-driven film studio , Claude MCP replacing trading bots , using OpenCore Legacy Patcher , Anthropic public agent harness , OpenAI reduced p95 latency , Claude Code 2.1.202 dynamic workflow size , Continue v1.2.24 shift , single line turning Claude into self-learning machine , agent definition with memory and tools , Claude Code generating game for $50 handheld , repo with 147 domain-specific skills , Commonly 2.0 group chat , single developer running 7 terminals , former quant replicating trading system , Claude Code connecting to Creatify via MCP , Claude transforming teacher workflow , treating Claude as thinking partner , running 20-30 coding agents concurrently , Cmux agentic terminal access , using Fable 5 for audit and agent team , most users missing features , company brain checklist , VideoAgent open-source , Anthropic free courses , Hermes judgment system , Codex /dynamic skill , vibecoding for niche tools , Claude building industry-specific agents , agent success depending on loop design , Anthropic workshop roadmap , Claude Code generating VST plugin , portable Raspberry Pi setup , LiteLLM unified interface , building three projects to get hired , ex-Google engineer building $50k marketing team , Claude executing trades on TradingView , Claude as thinking partner , deterministic rules with ast-grep , reusable Taste skill , structured Obsidian vault , Google Jules free agent , stack of memory and context tools , AI agent generating Fortnite maps , ruflo multi-agent system , Claude Code loop and goal features , physical gear shifter for model switching , Claude Code building trading bot , E Ink tablets for focused development , multi-agent dashboard for Congress , org chart structure for coordination , building system to generate prompts , Ghost MCP for disposable databases , Fable and Claude with session summaries , 165 specialized subagents , running local models for routine tasks , turning chatbot into coworker with Brain File , Claude agent for permit data lead generation , iterative prompt refinement for 3D models , structured system of context and skills , Claude Code modifying trading bot , Printing Press CLI , Claude Code pulling from Mobbin MCP , company knowledge base as graph , 5.3-Codex-Spark for UI tests , library of 75 design skills , Claude Code 2.1.204 ripgrep default , well-structured prompt making cheaper models perform , tablet LiDAR scanner with Claude Code , Fable 5 for finding tasks and refactoring , Claude building coffee shop website , Alook org chart tool , Claude building AI agent for listing videos , structuring Claude Code configuration , letting LLM build persistent wiki , Zed IDE free Claude access , using Fable 5 for planning and Codex for execution , letting AI extract strategy , structured prompts for exploration , orchestrator loop increasing success rate , Fable 5 for incident response , local MCP server for ComfyUI , Claude Code automating Reddit scraping , GEO-SEO Claude skill , 35-agent Fable research swarm consuming plan , Fable and Opus building 3D keyboard platform , moving Claude Code extension in Cursor .
Added 31 new corroborated practices (e.g., Claude Code for trading bots, video editing, website generation) and expanded tooling with Hermes, iFixAi, Freellmapi, and multi-agent orchestration tools. Updated contested section with many emerging practices from new claims.