What are the current best practices for coding agents?

Current best practices for coding agents center on agentic loops that self-correct and run autonomously, structured context engineering (CLAUDE.md, knowledge graphs), specialized skills over general prompts, and rigorous cost control. Strongly corroborated practices include using loop design over model selection, delegating model roles, and employing verification agents for 24/7 operation. The tooling landscape has expanded with open-source MCP servers, persistent memory, and multi-agent orchestration. Many emerging claims—massive agent parallelism, one-line self-improvement—lack broad validation.[8]

1. Design agentic loops (plan → act → verify → repeat) that self-correct; loop quality matters more than model choice ^{262,441,568,112,518}.
2. Build context with memory, instructions, examples, files, previous actions, tool results, and state for 8× productivity ¹⁸².
3. Add a CLAUDE.md file documenting project conventions to prevent over‑engineering and hallucinated APIs ^178,576.
4. Delegate roles across models: Fable for design, Opus for reasoning, Sonnet for execution ¹²⁵.
5. Use reusable skills instead of prompting from scratch; leverage skill libraries ^179,183.
6. Install the Find Skills skill to discover appropriate skills for any task ¹⁸³.
7. Use Codex’s /dynamic skill to spawn sub‑agents executing tasks in parallel ⁵⁷¹.
8. Employ a Plan‑Build‑Judge loop with three agents for iterative improvement ³¹⁷.
9. Run multiple Claude Code sessions in TMux with Git worktree isolation for adversarial reviews and parallel testing ³⁷⁶.
10. Use structured slash commands (/init, /memory, /agents) and a config file hierarchy for efficient project management ^383,290,858.
11. Leverage Anthropic’s official prompt library for tested prompts across the software lifecycle ¹⁷⁶.
12. Give agents a “Taste skill” defining UI preferences to avoid generic designs ⁵⁴¹.
13. Reduce system prompt size and use blind‑spot passes to unlock Claude Code’s full capability ⁸⁶¹.
14. Use separate verification agents to evaluate output against success criteria for 24/7 autonomous operation ¹⁰⁰⁶.
15. Write a HANDOFF.md at session end to preserve context and enable project continuity ⁹³⁶.
16. Implement a memory architecture with Obsidian for long‑term context, Hermes for active state, and routed RAG for cash‑generating performance ⁹³⁵.
17. Optimize costs by running local models for routine tasks and reserving cloud models for complex reasoning; convert code/JSON/logs to PNGs to slash token usage ^480,230.
18. Build agent businesses by first doing the work manually, then automating the learned process ³⁸¹.
19. Let Hermes learn repetitive workflows from a single demonstration and execute autonomously ³²¹.
20. Use two AI agents to cross‑verify each other’s work for higher accuracy ⁴⁴⁵.[28]

Core coding agents are Claude Code, Codex (parallel sub‑agent execution), and Hermes (workflow learning). Freellmapi aggregates free tiers of 16 LLM providers behind one local API. ai‑website‑cloner‑template reverse‑engineers live websites. Atomic Mail provisions agent email accounts in ~30 seconds. Claude‑Mem provides persistent memory cutting token use by 95%. Graphify creates knowledge graphs from codebases reducing token consumption 71.5×. OfficeCLI lets agents control Word/Excel/PowerPoint with one command. For multi‑agent orchestration: TMux + Git worktree, Cmux, Commonly 2.0, ruflo, and CNVS offer visual and organizational control. Anthropic’s public agent harness (277 skills, 67 agents, 34 rules, 15 hooks) enables 6‑hour autonomous runs and cross‑tool portability. Mobbin MCP provides 600,000 real UI screens for design reference.[16]

Many emerging practices lack broad validation and should be tested cautiously. Examples include: a single line of markdown making Claude a self‑learning work machine ⁶²¹, running 20–30 coding agents concurrently to ship 20–40 PRs per day ⁶¹¹, fully automating client prospecting ⁴⁴⁴, building a self‑prompting system that reduces token usage by 90% ⁸⁹², and using a 35‑agent Fable research swarm that consumes 50% of a $200 plan in 20 minutes ⁶⁹⁶. As the field evolves, some may become standard, but currently they are single‑source or anecdotal.[5]