What are the current best practices for coding agents?

Current best practices for coding agents emphasize 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]

1. Add a CLAUDE.md file documenting project conventions, rules, and context to prevent over-engineering and hallucinated APIs ¹⁷⁸.
2. Build context with 7 components—memory, instructions, examples, files, previous actions, tool results, state—to increase Claude Code productivity 8× ¹⁸².
3. 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.
4. Start with predefined workflows and add autonomy only when the task demands it ²²⁷.
5. Use Claude Code to automate video editing—jump cuts, captions, chapter titles, etc. ²²⁰.
6. Install the Find Skills skill to discover and apply the best skills for any goal ¹⁸³.
7. Use OpenAI’s codex-plugin-cc to let Claude Code delegate tasks to Codex in the same terminal ¹⁴¹.
8. Delegate roles across models: Fable for design, Opus for heavy reasoning, Sonnet for execution to optimize cost and performance ¹²⁵.
9. Use loops instead of single prompts to automate work ¹¹².
10. Point Claude Code at an Obsidian vault to turn notes into a searchable wiki ¹¹¹.
11. Use Freellmapi to aggregate free tiers of 16 LLM providers behind one local API ¹¹⁰.
12. Run iFixAi to red-team agent configurations with 45 tests and independent grading ¹⁰⁵.
13. Rely on Claude Fable 5 to generate complex interactive visualizations from code without external assets ¹⁰³.
14. Use the ai-website-cloner template to reverse-engineer live websites into clean code ⁴²⁸.
15. Build agent businesses by first doing the work manually, then automating the learned process ³⁸¹.
16. Run multiple Claude Code sessions in TMux with agent teams and Git worktree isolation for adversarial code reviews and parallel feature testing ³⁷⁶.
17. Optimize Hermes agent costs by model choice, background tasks, and context usage ³⁷².
18. Let Hermes Agent learn repetitive workflows from a single demonstration and execute them autonomously ³²¹.
19. Set up a pixel office: turn each Claude Code session into a visual character with speech bubbles, file tracking, and permission visualization ²⁶⁶.
20. Have Claude Code read and modify an open-source trading bot’s code to create a profitable automated strategy ⁴⁴⁶.
21. Use TouchDesigner, MediaPipe, and Claude Code to build real-time hand-controlled 3D applications ¹⁸⁰.
22. Copy-paste from Anthropic’s official prompt library for task-specific prompts ¹⁷⁶.
23. Use Claude Code with Sonnet 5 to produce a professional website in 18 minutes ¹³².[24]

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). 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. Many tools are early and lack extensive corroboration.[24]

Many practices remain based on single reports without independent confirmation. Examples: building a practical AI agent in under 10 minutes ²²⁴; generating a complete browser game from one spec file ⁴²⁹; achieving 624,900% margin on ad generation ¹¹⁶; shipping code 5× faster with Codex subagents ³⁸⁴; reducing engineering busywork by 50% ²⁹³; Anthropic running 99% of engineers with 300+ self-improving agent swarms ¹²⁹; a solar-powered mesh node serving a 3B model offline ¹⁷⁵; a $400 GPU rig replacing all cloud subscriptions ¹²⁷; Claude Code autonomously running a full Instagram content strategy ⁴⁴². Claims about specific performance lifts and niche hardware setups need broader validation before adoption.[9]