Job-by-job guides for Claude Code, Codex CLI, OpenCode, goose, Aider and Gemini CLI. Each one says which harness fits, why, and gives you config you can paste.
New recipes and harness changes, a few times a month.
# rig: literature review job cited brief from a folder of papers fits Claude Code · goose · OpenCode rig AGENTS.md citation rules .claude/agents/ reader.md one source → notes settings.json deny Edit(sources/**) run claude "review sources/txt/" docs linked + checked Oct 18, 2026
Comparisons and setup guides, sourced and dated.
The best AI coding agent harnesses of 2026 ranked on Artificial Analysis's Coding Agent Index (data as of Oct 18, 2026), with cost per task, licenses, and which harness fits which job..
OpenCode vs Claude Code compared on what matters for choosing: which models each runs, how you pay, config files, MCP, headless mode and the Artificial Analysis index.
Which agent harness to run on a local model with Ollama or LM Studio, with the exact config for OpenCode, Codex CLI, Aider and goose.
Which instructions file Claude Code, Codex, OpenCode 1 and 2, goose, Gemini CLI and Aider read, in what order, and a one-file setup that works in all of them.
Aider vs OpenCode for real work: git auto-commits and undo, instructions files, MCP, local models and scripting, with a shared setup that works in both.
The exact command and config file to add the same MCP server in Claude Code, Codex CLI, OpenCode 1 and 2, goose and Gemini CLI, with scopes and gotchas.
The harnesses that fit a real task, with the files to run each.
A copy-ready research rig for literature reviews: Claude Code subagents per source, a repeatable goose recipe, or OpenCode on a local model for unpublished papers.
A setup for writing books, reports and long docs with an agent harness: one file per chapter, a style guide every agent loads, per-change git commits in Aider, and a read-only editor subagent in Claude Code..
The exact headless commands for running Codex, Claude Code, OpenCode, goose, Gemini CLI and Aider in CI or cron jobs, with the least-privilege flags and JSON output for each.
A read-only, local-model rig for questions over private files: OpenCode, goose or Codex on Ollama, with web access and edits locked down.
New to the idea? Read coding agent harness, explained.
Six one-card setups: a harness, a job and the config surface that makes it work.
Give each source its own subagent context, then have the main session merge the notes into one cited brief. A hook can block edits outside an output folder.
Point OpenCode at Ollama's OpenAI-compatible endpoint and set permissions to ask before edits or shell commands. Also set small_model to a local model so lightweight tasks stay local too.
codex exec runs without the TUI and prints the final message to stdout. It is read-only by default, so allow edits explicitly with the workspace-write sandbox. --json streams JSONL events for scripts.
A goose recipe is a YAML file with instructions, a prompt and parameters. Run it the same way each time and pass new parameter values per run.
Aider commits each change it makes with a descriptive message. /diff shows what changed and /undo discards the last change, so you can review and roll back one step at a time.
With OpenHands Cloud connected to your repo, label an issue openhands (or comment starting with @openhands). It comments while it works and opens a pull request if it decides the issue is resolved.
Leaderboards rank harness + model pairs. We start from the task and name the 2–3 harnesses that fit it.
AGENTS.md, CLAUDE.md, MCP servers, permission rules and headless commands, in each harness's own syntax.
Every fact links to official docs or the cited leaderboard, with the date we checked. Unknowns are marked, never guessed.