Run any agent harness natively.
Your team already chose a harness - the runtime whose behavior is your agent's behavior. Karta runs Claude Code, OpenCode, DeepAgents, Goose, and Codex CLI in their native project formats. No translation into a proprietary workflow graph, ever.
The harness is the application boundary.
Your folder is the deployable
The same project you develop locally - skills, tools, subagents, harness config - deploys as-is. The harness is auto-detected from the folder; pin it explicitly only when you want to.
Versions are pinned facts
Every release freezes its resolved harness and version. A harness upgrade is a decision you make and can roll back, not a surprise you discover in production.
Readiness-gated, security-held
A version reaches "supported" by passing real deploys and turns; versions with security advisories are pulled until a fixed release certifies. The matrix is public.
Uniform behavior across all five
Sessions, streaming, approvals, and traces behave identically whichever harness runs underneath - one integration on your side, five runtimes on ours.
Switching harnesses later is the point, not a migration: your agents' workspaces, users, and history survive the swap. Building a harness of your own? Talk to us - running in-house harnesses on the same contract is in early access.