Overview

The AI agent tooling stack splits into three jobs: write code (coding agents), build agents (frameworks), and trust agents (evaluation and observability). This wiki compiles what the curated sources actually say about how these tools differ, where they overlap, and how to choose between them.

The recurring themes, each with its own page:

• How agents keep context manageable without drowning the model: context management.
• The freedom to bring your own model, including local ones: BYOK and model choice.
• The Model Context Protocol as a shared tool interface: MCP.
• Running tooling on your own infrastructure for privacy: self-hosting.
• Git as the audit log and rollback mechanism: git-native workflows.
• Telling tracing apart from evaluation: evaluation and observability.
• The cognitive cost of agentic coding: comprehension debt and skill atrophy.
• Where the real leverage lives: agent leverage and loops.
• Why context as the bottleneck moves the limit to specs and shared context.
• Patterns for code review with agents: self-improving and multi-lens review.
• Designing codebases agents work well in: agent experience.
• The highest-leverage habit: specs and planning.
• Treating LLM calls as structured input and output.

Entity pages go deep on individual tools, starting with Claude Code, Aider, OpenCode, and Langfuse.

The directory at /tools has the verified specs; this wiki is the practitioner-readable synthesis on top.

Sources: 0001-claude-code-2000h.md, 0002-aider-docs.md, 0003-opencode-docs.md, 0004-langfuse-docs.md, 0007-karpathy-claude-workflow.md, 0008-agentic-coding-trap.md, 0009-bottleneck-never-code.md, 0010-self-improving-code-review.md, 0011-multi-agent-review-opencode.md, 0012-addyosmani-llm-workflow.md, 0013-marmelab-agent-experience.md, 0014-mamdani-mcp-guide.md, 0015-shuaiguo-ten-lessons.md

Edit wiki/overview.md; see how we maintain the wiki.