Kodo

What it claims

Kodo adds an orchestration layer over coding agents so a goal set at night yields tested, reviewed code by morning, running workers on a Claude Code Max subscription while an orchestrator delegates. The headline number the summary cites gets a critical qualifier here: the 24% gain was a 100-task head-to-head using the same underlying model (Cursor composer-1.5) — same model, same prompt, same conditions — isolating orchestration as the sole variable, not comparing a strong model against a weak one.

How it works

Beyond the orchestrator→team structure the summary describes, the page documents a verification loop as the core mechanism: workers done(), then tester, tester_browser, and architect each independently sign off, and any rejection sends work back. The showcased blackopt run went 9 rejection rounds (resource leak, class-variable contamination, state mutation, crossover edge cases) before acceptance — the architect verifier caught bugs the worker was blind to.

The page also argues for an API orchestrator over a CLI one: CLI coding tools (Claude Code, Cursor, Codex) try to write code themselves and micromanage, whereas a plain API model stays in a delegation-only lane.

Two modes the summary omits entirely:

• kodo test — exercises the software as a real user (install, every feature/flag, edge cases), writes failing regression tests, then fixes. Tracks coverage in .kodo/test-coverage.md.
• kodo --improve — senior-dev code review for simplification/usability/architecture, auto-fixes safe issues, flags ambiguous ones.

How to use it

uv tool install kodo-agent
uv tool install 'kodo-agent[benchmark]'   # adds kodo-bench SWE-bench harness

Needs uv and at least one backend; Claude Code plus one fast backend (Cursor/Codex/Gemini CLI) recommended. Set GOOGLE_API_KEY (Gemini orchestrator, "fast and cheap") or ANTHROPIC_API_KEY.

kodo --goal 'Build X' --team full --exchanges 50 --cycles 10 ./my-project
kodo --resume 20260218_205503
kodo --fix-from <RUN_ID>

Flags worth knowing: --effort low|standard|high|max, --auto-refine and --skip-intake for unattended runs, --no-auto-commit, --json. Cost is split into Virtual (subscription-covered, workers) and API (real spend, orchestrator).

Limits

The page's own warning: agents run in bypassPermissions mode with full permissions — they primarily work in the project dir but can access any file on your system (install deps, edit configs). It explicitly says to have a git commit or backup before launching. Browser/Docker/VPS tasks that agents can't self-provision surface under a "Blocked Workflows" report section rather than being handled.