Agents that stay in their lane.
AI Agents (Helios Core Six) is a task engine with one owner per task, hard role boundaries, and autonomy that has to be earned.
The moment more than one agent does real work, they duplicate each other and drift outside their remit. That is designed out here with locks and boundaries rather than asked for politely in a prompt.
This is a deployment definition rather than a packaged product today. Treat it as the least mature of the three.
How work is controlled
- Task engine
- Every task is a row with a goal, a definition of done and exactly one owner, moving through a defined state machine.
- Atomic claim-lock
- Exactly one agent can take an open task. Two cannot pick up the same work.
- Structured handoffs
- Defined transfers between agents rather than free-form chat.
- Role boundaries
- Each agent has one job and a hard boundary on the tools it can reach.
- Earned autonomy
- Three rungs, from draft-only, to act-then-review, to autonomous. Promotion and demotion are mechanical, driven by a track record rather than a switch.
- Human approval gate
- Any action that changes the outside world passes a human gate at that action's autonomy level. Payments and destructive changes never earn autonomy at all.
- Loop and injection guards
- Agent-to-agent reply loops are capped. A detected prompt injection freezes the affected category to draft-only.
- Per-agent scoped credentialsNot yet available
- Each agent scoped to its own tokens rather than a shared set.