honest comparison. we win where it matters.

a live, navigable architecture graph of your repo — crates, packages, entry points, import edges. the agent reads it too.

durable, in-repo memory the agent can read and write with vault_read / vault_write. not a rules file — a real knowledge store.

9 step kinds, bundled seeds, long-running spawned agents. compose multi-step engineering pipelines, not just one-shot prompts.

write, ask, plan, debug, secure, minimal, architect, refactor, performance — each with its own tool allowlist and default model.

every tool call written to a portable sqlite file on your disk. real auditability without a cloud dashboard.

open agent control protocol with external agent servers and an in-app marketplace for qa, security, and perf agents.

first-class isolation — every agent run lives in its own worktree, not a shared branch. no stomping on your work.

per-message routing chooses the right model for the task. you stop picking models; we pick the cheapest one that wins.

ship agent changes through a real evaluation harness, and diagnose any config in one slash command.