Agentic Coding

This content directly addresses all three of your search intents. The repository documents Claude Code workflows, agentic coding practices, and context window management strategies specifically tailored for solo founders and SaaS developers. For workflow orchestration, the repository presents the Command → Agent → Skill architectural pattern with a complete /weather-orchestrator example. It covers dynamic workflows, agent teams with parallel compute using git worktrees, and systematic development workflows following Research → Plan → Execute → Review → Ship phases across multiple curated repositories. For agentic coding and debugging strategies, it provides extensive best practices including: using subagents for context isolation (keeping intermediate tool calls out of main context), cross-model workflows combining Claude Code with other models for QA-review loops, test-time compute where separate agents find bugs that others introduce, and specific debugging techniques like taking screenshots, using Chrome DevTools MCP, and running terminal logs as background tasks. The repository emphasizes feature-specific sub-agents over generalist ones and recommends using /code-review for multi-agent PR analysis. For context window management, it details practical strategies: context rot occurs around 300-400k tokens on the 1M model; the "dumb zone" kicks in at 40% context usage (keep under 40%, aggressive users target 30%); rewinding is preferred over corrections to avoid polluting context; manual /compact with hints beats autocompact; and subagents should be used strategically to keep child context isolated while returning only conclusions to the parent. The repository includes specific tips on session management, when to use /clear versus /compact, and using recaps for long-running sessions.

The ECC (Enterprise Claude Code) system addresses all three of your core interests directly: For Claude Code workflow orchestration, ECC provides a production-ready agent harness operating system (v2.0.0) with 64 specialized subagents, 262 reusable skills, orchestration commands (/multi-plan, /multi-execute, /multi-workflow), and PM2 multi-service lifecycle management. The subagent orchestration skill explicitly handles the context problem through iterative retrieval patterns and cascade methods, with session adapters for structured recording and Rust-based control-plane infrastructure for deterministic session management. On agentic coding debugging, ECC ships specialized resolver agents for build errors across 12 language ecosystems (TypeScript, Python, Go, Java, Kotlin, Rust, C++, PHP, Perl, Swift, F#, HarmonyOS), plus verification loops, checkpoint management, and continuous learning v2 (instinct-based pattern extraction with confidence scoring). The hook runtime includes deterministic harness audit scoring, observer loop prevention with 5-layer guards, and a complete evaluation harness with pass@k metrics and grader types for verification-driven workflows. For context window management, ECC provides token optimization rules (model selection, system prompt slimming, background process tuning), memory persistence hooks that auto-save/load context across sessions, SessionStart context capping (8000 characters default, configurable down to 4000 or disabled entirely), and explicit MCP inventory control to prevent context explosion. The system ships runtime controls (ECC_SESSION_START_MAX_CHARS, ECC_SESSION_START_CONTEXT=off for low-context setups) and cross-harness memory isolation (separate ~/.cursor/ecc vs ~/.claude paths) so you can tune strictly per tool without mixing state.

Claude Code's Agent Teams enable coordinated multi-agent orchestration with dedicated context windows per task, shared task lists with dependency tracking, and direct inter-agent messaging—differentiating it from Codex's isolated manager-worker model. Claude Opus 4.7 now leads on SWE-bench Pro (64.3% vs 58.6%) and supports a 1M token context window versus Codex's 200K, though Claude uses approximately 4x more tokens per task due to more thorough reasoning. Both tools now ship production-ready subagent support: Codex allows up to 8 parallel workers in cloud sandboxes, while Claude Code has no hard agent cap but burns limits proportionally per spawned agent. For agentic coding workflows, Claude Code excels at complex architectural refactoring and long-running autonomous tasks through Agent Teams with /goal for persistent multi-day objectives and /ultrareview for parallel code review. Codex optimizes for speed and spec-driven parallel execution through its manager-worker pattern and Memories MCP server for cross-session context. Token consumption differences matter significantly for cost: using Claude Sonnet 4.6 for worker agents and Opus 4.7 only for lead agents can substantially reduce expenses on multi-agent workloads. Both platforms now include effort levels, persistent goals, cross-session memories, and plugin ecosystems with hooks, with Claude Code accounting for approximately 10% of all public GitHub commits (~326K daily as of June 2026).

Antigravity Awesome Skills is an installable library of 1,527+ reusable SKILL.md playbooks designed for Claude Code, Cursor, Codex CLI, Gemini CLI, and other AI coding assistants. The library provides structured operating instructions to help agents perform recurring tasks with better context, stronger constraints, and clearer outputs—directly addressing the user's interest in efficient agentic coding workflows and advanced agent orchestration. The repository includes specialized plugins for focused domains (web apps, security, DevOps, QA, product design, data analytics, agent/MCP building), bundles for role-based skill discovery, and workflows for ordered execution of planning, coding, debugging, testing, and deployment tasks. For context window management, the library offers agent-overload-recovery guidance and selective installation options using --risk, --category, and --tags flags to keep the runtime skill set narrow and avoid context limits. Users can install via npx antigravity-awesome-skills with tool-specific flags for Claude Code, Cursor, Codex, or other platforms, and access curated starting points like the SaaS MVP bundle combining Essentials, Full-Stack Developer, and QA & Testing skills.