Human-AI Teaming¶
G6 is designed for collaboration, not autonomy. Here's why human judgment matters.
The Principle¶
AI systems improve faster with human feedback than with self-evaluation alone. G6 is built around this principle — it proposes, you approve.
How Collaboration Works¶
1. Safety-Critical Operations¶
When G6 encounters a high-risk operation (code execution, file modification, external API calls), it pauses for human approval via the HITL (Human-in-the-Loop) system.
2. Skill Evolution Review¶
The EvoSkill evolution loop can run autonomously, but production deployments should enable human review of proposed skills before they enter the frontier.
3. Goal Decomposition Validation¶
When G6 decomposes a complex goal, it presents the subtask tree for review. You can modify priorities, add constraints, or reject subtasks.
The HITL Approval Workflow¶
G6's HITL system implements a blocking approval gate for safety-critical operations. When an operation exceeds the CSF safety threshold, the pipeline pauses and creates an approval request.
Approval surfaces¶
Approval requests can be reviewed through:
- TUI Safety tab — visual approval/reject with evidence trail
- REST API — 11 dedicated HITL endpoints (
/api/hitl/pending,/api/hitl/approve/{id},/api/hitl/reject/{id}, etc.) - MCP tools —
tier_confirm_advanceandtier_reject_advancefor tier advancement gates
How the gate works¶
graph TD
A[Operation requested] --> B{CSF safety check}
B -->|Below threshold| C[Execute immediately]
B -->|Above threshold| D[Create HITL approval request]
D --> E{Human reviews}
E -->|Approve| F[Execute with audit trail]
E -->|Reject| G[Return rejection to caller]
E -->|Timeout| H[Fail safe — operation blocked]Breakpoints and guardrails¶
GoalInput supports three levels of human oversight:
| Control | Behaviour | Use case |
|---|---|---|
| Guardrails | Hard constraints — operation fails if violated | Budget limits, forbidden operations |
| Checkpoints | Soft assertions — logged but non-blocking | Quality thresholds, expected ranges |
| Breakpoints | Human pause points — pipeline halts until approved | Review intermediate results, validate direction |
Advisory vs Enforcing modes¶
Safety gates operate in two modes:
- Advisory (default for low-risk operations): the gate logs a warning but allows execution
- Enforcing (default for high-risk operations): the gate blocks execution until approved
The mode is configurable per component via the CSF configuration.
When to Trust G6¶
| Operation | Trust Level | Default Mode |
|---|---|---|
| Information retrieval | High | Advisory |
| Data analysis | High | Advisory |
| Code generation | Medium | Advisory (review output) |
| Skill evolution | Medium | Enforcing (review proposals) |
| File modification | Low | Enforcing |
| External API calls | Low | Enforcing |
| Tier advancement | Low | Enforcing (requires HITL confirm) |
FAQ¶
Q: Can I run G6 fully autonomously? A: Yes, for low-risk operations. High-risk operations always require approval unless explicitly disabled via CSF configuration.
Q: How do I review evolution proposals? A: Use the list_programs and get_frontier tools to inspect what G6 has discovered. Use tier_confirm_advance or tier_reject_advance for tier advancement decisions.
Q: What happens if I don't respond to an approval request? A: The operation times out and fails safe — no action is taken. The timeout is configurable via HITL_TIMEOUT_SEC (default: 300 seconds).