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. Key practices include using CLAUDE.md files to document project conventions ¹⁷⁸, building context with 7 components to boost productivity 8x ¹⁸², and employing agentic loops (plan → act → verify) for self-correction ^262,317. The field is rapidly evolving with emerging tools like ECC ¹⁷⁷ and Hermes ³²¹ that enable multi-agent orchestration. However, many claims are based on single reports and lack broad corroboration.[6]

1. Use a CLAUDE.md file to document project conventions, rules, and context to prevent common errors like over-engineering and hallucinating APIs ¹⁷⁸.
2. Build context with 7 components—memory, instructions, examples, files, previous actions, tool results, state—to increase Claude Code productivity 8x ¹⁸².
3. Employ agentic loops (plan → act → verify → repeat) for self-correction and iterative improvement; Anthropic uses this loop to write over 40% of code ²⁶².
4. Use structured slash commands and workflows (e.g., /init, voice mode, planning, verification) to improve productivity ¹²¹.
5. Provide a design system before generating screens to ensure consistent visual identity ²²⁵.
6. Use the 'handoff' skill to compress session context into a markdown file, allowing fresh agents to continue without degradation from long context windows ²²⁶.
7. Start simple with predefined workflows and add autonomy only when the task demands it ²²⁷.
8. Use Fable for overall design, Opus for heavy reasoning, and 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 multiple AI agents (Claude and DeepSeek) in a single conversation for drafting, challenging, and refining to improve output quality ¹³³.
12. Use the Find Skills skill to discover and apply the best skills for any goal ¹⁸³.
13. Use Claude Code to automate video editing including jump cuts, captions, chapter titles, sound effects, BGM, logo, B-roll, and thumbnail generation ²²⁰.
14. Use Claude Code with Sonnet 5 to produce a professional website from design to code in 18 minutes ¹³².
15. Use Claude Fable 5 to generate complex interactive visualizations entirely from code without external assets ¹⁰³.
16. Use iFixAi for automated red-teaming of agent configurations ¹⁰⁵.
17. Use Freellmapi to aggregate free tiers of 16 LLM providers behind one local API to avoid paid caps ¹¹⁰.
18. Use OpenAI's codex-plugin-cc to let Claude Code delegate tasks to Codex in the same terminal for multi-agent collaboration ¹⁴¹.
19. Use Hermes Agent to learn and autonomously execute repetitive workflows after a single demonstration ³²¹.
20. Use a pixel office setup to turn each Claude Code session into a visual character with speech bubbles, live file tracking, and permission visualization ²⁶⁶.
21. Use Claude Code to read, understand, and modify an open-source trading bot's code to create a profitable automated strategy ⁴⁴⁶.[21]

Key tools include: Claude Code (primary agent), ECC (open-source config with 60 agents and 231 skills) ¹⁷⁷, Hermes Agent Dashboard (supports custom plugins) ¹²⁸, Codex OAuth (secure multi-agent orchestration) ³²⁰, Browser Use CLI 3.0 (6x token reduction) ²¹⁷, and Fable (one-shot CLI coding agent) ¹³⁵. For cost reduction: Freellmapi aggregates free tiers of 16 LLM providers ¹¹⁰, and converting code to PNG images reduces context token usage by ~80% ²³⁰. For multi-agent collaboration: OpenAI's codex-plugin-cc lets Claude Code delegate tasks to Codex ¹⁴¹, and Orca organizes agents into parallel teams ¹³⁷. For security: iFixAi provides automated red-teaming ¹⁰⁵. For local inference: a $400 rig of four GTX 1080 GPUs running Ollama can replace cloud subscriptions ¹²⁷. Other notable tools: Dotcode (browser-native agentic environment) ¹³⁰, Morph (200 free AI requests/month) ¹²⁴, Agent-Reach (single command to access 14 platforms) ¹¹⁴, Lev8 (lead generation with 90 valid results) ¹¹⁷, and CNVS (visual orchestration of multiple agents) ³²².[16]

Many claims are based on single reports and lack broad corroboration. For example: building a practical AI agent in under 10 minutes ²²⁴, generating a complete browser game from a single spec file ⁴²⁹, and achieving 624,900% margin on ad generation ¹¹⁶ are unverified. Claims about specific performance numbers (e.g., 5× faster shipping with Codex subagent harnesses ³⁸⁴, 50% reduction in engineering busywork ²⁹³) need independent validation. The effectiveness of using Claude Fable for final code review ³¹⁴ and the claim that Claude Code power users achieve 'higher performance' ¹²¹ are vague. The claim that Anthropic runs 99% of engineers with 300+ self-improving agent swarms ¹²⁹ is unverified. Other unverified claims include: using Claude to monitor BTC every 5 minutes with 31 simulations ³¹⁹, a solar-powered mesh node running a 3B model ¹⁷⁵, and the claim that Claude Code can autonomously run a full content strategy loop for Instagram ⁴⁴². Users should treat these as emerging practices and test in their own contexts.[11]