Version-control your company with a workforce of agents

What it does

Kortix is an AI-native operating layer for organizations. It bundles your agents, skills, integrations, triggers, and memory into a single Git repository anchored by a kortix.toml manifest. Agents run inside disposable Linux sandboxes on isolated branches, committing work that reaches main only after a human approves the change request. The pitch is that your company self-improves one reviewed diff at a time, even while the office is empty.

The interesting bit

The architecture treats a company as literal code, not a metaphor. Every session spins up its own branch and sandbox; the agent can break anything inside it, but only what it commits survives. It is essentially GitOps for corporate memory and labor, with approval gates keeping the agents from rewriting reality unchecked.

Key highlights

• Agents, skills, connectors to 3,000+ apps, secrets, channels, and memory all live as code in one repo.
• Work runs on-demand, human-assisted, or fully automated via cron and signed webhooks.
• Self-hostable from laptop to air-gapped VPC; the same CLI toggles between managed cloud and local Docker hosts.
• Bring-your-own-model: provider-agnostic, with encrypted secrets injected at runtime and never exposed to logs or the model.
• Enterprise-grade RBAC, audit trails, and microVM isolation.

Caveats

• The repo description and first README call it “open source,” but the license is Elastic License 2.0 (source-available), a discrepancy the second README quietly fixes.
• The “living company brain” memory is plain files today; the compounding part is still on the roadmap.
• The coding-agent runtime is “becoming pluggable,” so engine choice is not fully realized yet.

Verdict

Worth a look if you are building an AI-native org and need auditable, versioned agent output instead of black-box chat. Skip it if you just want a better ChatGPT wrapper; this is infrastructure, not a conversation app.