What are the current best practices for coding agents?

Current best practices for coding agents center on structured context (CLAUDE.md, knowledge vaults), skill-based specialization (Claude Skills, Find Skills), and agentic loops that self-correct and run autonomously. Corroborated practices emphasize loop design over model selection, delegating roles across models for cost/performance, and using verification loops to eliminate bugs. The tooling landscape is dominated by Claude Code, with Hermes and Codex as powerful complements, while cost optimization via local models and token reduction is key. Many emerging practices lack broad validation and should be tested cautiously.[9]

1. Use agentic loops (plan → act → verify → repeat) for self-correction and autonomous work; loop design matters more than model choice ^{112,262,518,441,568}.
2. Build context with 7 components—memory, instructions, examples, files, previous actions, tool results, state—to boost Claude Code productivity 8× ¹⁸².
3. Add a CLAUDE.md file documenting project conventions, rules, and context to prevent over-engineering and hallucinated APIs ^178,576.
4. Delegate roles across models: Fable for design, Opus for heavy reasoning, Sonnet for execution to optimize cost and performance ¹²⁵.
5. Use loops instead of single prompts; treat coding as designing reusable prompt patterns ^107,112.
6. Employ a Plan-Build-Judge loop with three agents for self-correction and iterative improvement ³¹⁷.
7. Use a verification agent in autonomous loops to evaluate executor output against success criteria for 24/7 completion ¹⁰⁰⁶.
8. Run multiple Claude Code sessions in TMux with Git worktree isolation for adversarial code reviews and parallel testing ³⁷⁶.
9. Use structured slash commands (/init, /memory, /agents) to improve productivity across writing, coding, and analysis ^383,290.
10. Install the Find Skills skill to discover and apply the best skills for any goal ¹⁸³.
11. Leverage Anthropic’s official prompt library for copy-paste prompts across the software lifecycle ¹⁷⁶.
12. Use Codex’s /dynamic skill to spawn sub-agents executing tasks in parallel, mimicking Claude Code’s Dynamic Workflows ⁵⁷¹.
13. Use Fable 5 for planning and Codex for execution to improve coding agent effectiveness ⁷⁵⁴.
14. Optimize costs by running local models (Ollama, Qwen, Llama) for routine tasks and reserving Claude for complex reasoning, cutting costs by 80% ⁴⁸⁰; also convert code to PNG to cut token costs ~80% ²³⁰.
15. Build agent businesses by first doing the work manually, then automating the learned process ³⁸¹.
16. Let Hermes Agent learn repetitive workflows from a single demonstration and execute them autonomously ³²¹.
17. Give Claude agents a reusable “Taste skill” defining UI preferences to avoid generic interfaces ⁵⁴¹.
18. Reduce system prompt size and use blind-spot passes and brainstorming prototypes to unlock Claude Code’s capability ⁸⁶¹.
19. Follow a structured file hierarchy for Claude Code configuration to enable efficient solo development and reuse across projects ²⁹⁰.
20. Use a stack of agents, commands, settings, hooks, MCPs, and templates to turn Claude Code into a structured development team ²⁶³.[26]

The core coding agent is Claude Code, augmented by Hermes (learns workflows from one demo and can be cost-optimized) and Codex (for parallel sub-agent execution, OAuth permission-based delegation). Freellmapi aggregates free tiers of 16 LLM providers behind one local API to avoid paid caps. The ai-website-cloner-template enables reverse-engineering live websites into clean code. Atomic Mail provides email accounts for AI agents via API in ~30 seconds. Claude-Mem reduces token usage by 95% with persistent memory. For multi-agent orchestration, tools like TMux + Git worktree, Cmux, Commonly 2.0, ruflo, and CNVS enable visual and organizational control. Anthropic’s public agent harness (277 skills, 67 agents, 34 rules, 15 hooks) promises 6-hour autonomous runs and cross-tool portability. Many other specialized agents and MCP servers are emerging.[17]

While the playbook above relies on corroborated practices, many emerging claims have only single-source support and should be tested cautiously. For example, using AI to fully automate client prospecting ⁴⁴⁴, running 20–30 coding agents concurrently to ship 20–40 PRs per day ⁶¹¹, or using a single line of markdown to make Claude a self-learning work machine ⁶²¹ are unverified. The tooling landscape is also in flux, with new MCP servers, orchestration frameworks, and model releases appearing weekly. As the field matures, some of these emerging practices may become standard, but currently they lack the broad validation of the corroborated set.[3]