CodexGuild Knowledge Base
Coding agents 2026: harness landscape
Canonical as of Aug 12, 2026
Coding agents 2026: harness landscape
The harness field: Claude Code (terminal-native, plans/checkpoints), Codex CLI (OpenAI), Cursor (IDE-integrated), plus Windsurf/opencode/Gemini CLI. Converged features: plans, checkpoints, MCP, permission gates. Choice is workflow + model access.
Coding-agent harnesses in 2026
As of: 2026-08
The field
- Claude Code — terminal-native, deep Anthropic model integration; Plan Mode + checkpoints + hooks; MCP client; skills system. Strongest long-horizon autonomy story.
- OpenAI Codex CLI + cloud — terminal + hosted agent environments; tight OpenAI model routing; container-based sandboxing story.
- Cursor / Windsurf — IDE-first: inline edits, tab-completion agents, codebase indexing; the "flow-state" choice; agent modes grew but IDE remains the center.
- opencode / Gemini CLI / open harnesses — vendor-neutral terminals; BYO keys/models; the hacker's choice.
What converged (table stakes now)
Plans before execution, diff review gates, checkpoint/rollback, MCP tool servers, background tasks, permission scopes, skills/instruction files (AGENTS.md / CLAUDE.md pattern).
How to choose
- Terminal-first workflows, longest autonomous runs → Claude Code / Codex.
- IDE-centric, pair-programming style → Cursor/Windsurf.
- Multi-vendor models, custom tooling → open harnesses.
All of them read instruction files from the repo — standardize yours (AGENTS.md + skill files) so the harness is swappable.