Naturally Autonomous: Why Agents Should Organize Themselves

The name is the thesis

The name describes the intended result: several independent voices producing one coherent piece of work rather than four soloists. The thesis behind the machinery is natural autonomy, a claim about how multi-agent systems should work.

Most agent tooling today asks you to be the coordinator: pick the model for each task, wire the workflow, shuttle context between chats, babysit the handoffs. The agents are capable; the organization is manual. You end up as a project manager for software.

We think that's backwards. If agents are going to do real engineering, they should organize themselves, the way a good team does.

What natural autonomy looks like

Give a team of people a goal and they don't wait for you to assign every subtask. They divide the work along their strengths, ask each other for help, review each other's output, and merge the results. The coordination is a property of the team, not a burden on the client.

Harmony is built so agents can do the same:

• Each agent has its own role, skills, memory, and tools
• Agents message each other through durable envelopes: ask, reply, review, hand off
• Work is divided across a real structure: departments, leads, members, isolated worktrees
• Output is reviewed by other agents before it reaches you
• The user describes the goal; the system figures out the best way to accomplish it

Autonomy without opacity

The obvious objection: autonomous agents sound like a black box. That's why transparency is a pillar, not a feature. Every prompt, tool call, permission check, and handoff lands in a local trace ledger you can read. Spawning a new agent requires your explicit, single-use approval. Security-relevant actions go to a tamper-evident audit log.

Natural autonomy doesn't mean unsupervised. It means the supervision happens at the level of goals and evidence, not keystrokes.

Cooperative, not competitive

One more choice baked into the name: agents in Harmony collaborate rather than compete. They share knowledge, delegate along capability lines, and escalate to a human when a decision is genuinely yours to make. Different models, from Claude and Codex to Grok, Cursor, and open-weight models like Kimi K3, work as one coordinated system, because the coordination layer is provider-agnostic by design.

A single model acting alone is a tool. A team of specialists working toward a common objective is something more. That's the bet: cooperative intelligence, running on your machine, free, with your own models and agents.

See what this looks like in practice: several agents, one monorepo.