Examples

Approval-gated agent

Autonomy is what gates actions, and it lives in intent.toml:

# axl-config/agents/ops/intent.toml
[objective]
problem = "Handle routine operational requests"
why = "Reduce manual toil without giving up oversight"
success_criteria = ["Consequential actions are confirmed by a person"]

[constraints]
rules = ["Never modify production configuration without approval"]

[autonomy]
level = "require_approval"

[stop_rules]
rules = ["Stop when the request is complete or approval is refused"]

To be asked about risky calls rather than every call, add the risk gate:

# axl-config/agents/ops/agent.toml
[middleware.risk_gate]
enabled = true

And cap how much a single confused run can change:

[middleware.blast_radius]
enabled = true
max_writes = 20

Approval in AXL is driven by autonomy, not by whether a tool is a write.

Next

Agent with memory

Preferences and facts that follow a user across conversations.

Cost-controlled agent

Caching, routing, and caps working together on one agent.

On this page

Next