Overview
Docker Agent has Safety Modes and single Evaluation Order.
I need an autonomous safe (pre-configured) agent which works on tasks for hours without my (human) engagement. Let's call it Human-On-The-Loop (not Human-In-The-Loop).
Let mean under "an autonomous agent" such an agent which has all appropriate permissions and no asks human's confirmation.
Let mean under "a safe agent" such an agent which constrains to perform only "safe operations" (safety mode "strict" or "balanced") plus operations which user marks as safe ones.
However, right now I have to choose:
- use
balanced safety mode to automatically allow safe operations BUT manually confirm destructive AND unknown ones
- use
autonomous safety mode to automatically allow all operations (destructive AND unknown)
But actually, my needs are:
- autonomous working (no manual human confirmation)
- allow my custom operations which I describe via toolset (like lint, test-unit, test-e2e)
- allow safe operations (in terms of "safety modes")
- deny everything else
Motivation
I want to create a safe autonomous agent easily.
I expect my agent will works for the hours without human's confirmation on operations.
I expect reusing the built-in "safety patterns".
Use cases
No response
Proposed solution
Custom evaluation steps and order
Let's imagine I can configure evaluation order:
deny (by default) -> allow -> ask (it's current implementation)deny (by default) -> allow (no ask everything, it's very close to my needs)allow (safe ops + white list via permissions) -> deny everything else (it is exactly what I need)
Switching decision types for safety categories
Let's imagine I can configure decision type for current safety categories to build my own rules:
allow:safedeny:destructivedeny:uknown
Power of both: Custom evaluation steps and order + Switching decision types for safety categories
Now, I can easily describe what I need:
- Set the order:
allow -> deny (without ask step):
Allow
- Add allowed permissions for my custom tools like lint, test-unit, etc
- Reuse safety patterns for other safe operations
Deny everything else
Alternatives
No response
Related issues
No response
Additional context
No response
Overview
Docker Agent has Safety Modes and single Evaluation Order.
I need an autonomous safe (pre-configured) agent which works on tasks for hours without my (human) engagement. Let's call it Human-On-The-Loop (not Human-In-The-Loop).
Let mean under "an autonomous agent" such an agent which has all appropriate permissions and no asks human's confirmation.
Let mean under "a safe agent" such an agent which constrains to perform only "safe operations" (safety mode "strict" or "balanced") plus operations which user marks as safe ones.
However, right now I have to choose:
balancedsafety mode to automatically allow safe operations BUT manually confirm destructive AND unknown onesautonomoussafety mode to automatically allow all operations (destructive AND unknown)But actually, my needs are:
Motivation
I want to create a safe autonomous agent easily.
I expect my agent will works for the hours without human's confirmation on operations.
I expect reusing the built-in "safety patterns".
Use cases
No response
Proposed solution
Custom evaluation steps and order
Let's imagine I can configure evaluation order:
deny(by default) ->allow->ask(it's current implementation)deny(by default) ->allow(no ask everything, it's very close to my needs)allow(safe ops + white list via permissions) ->denyeverything else (it is exactly what I need)Switching decision types for safety categories
Let's imagine I can configure decision type for current safety categories to build my own rules:
allow:safedeny:destructivedeny:uknownPower of both: Custom evaluation steps and order + Switching decision types for safety categories
Now, I can easily describe what I need:
allow->deny(without ask step):AllowDenyeverything elseAlternatives
No response
Related issues
No response
Additional context
No response