ulw

Ultra Light Work — autonomous mode where the agent keeps working turn after turn without approval prompts, until it reaches a checkpoint.

Tell the agent to "just do it" — no approval prompts, no interruptions. Agent works until done or until the turn limit.

Usage

main.py

from connectonion import Agent
from connectonion.useful_plugins import tool_approval, ulw

# ulw requires tool_approval to work
agent = Agent("worker", plugins=[tool_approval, ulw], model="co/claude-opus-4-5")

ULW mode works by setting a skip_tool_approval flag that tells the tool_approval plugin to skip all checks.

Triggering ULW Mode

From the web UI

In co ai, use the mode selector to switch to ULW mode. The frontend sends:

code

{ "type": "mode_change", "mode": "ulw", "turns": 10 }

Programmatically

main.py

from connectonion.useful_plugins.ulw import handle_ulw_mode_change

handle_ulw_mode_change(agent, turns=10)
agent.input("Refactor all Python files to use type annotations")

Turn Checkpoints

When the agent reaches ulw_turns, it pauses and the frontend receives:

code

{ "type": "ulw_turns_reached", "turns_used": 10, "max_turns": 10 }

The user can respond with:

{ "action": "continue", "turns": 10 }

Extend by N more turns and keep going

{ "action": "switch_mode", "mode": "safe" }

Return to safe mode (approval prompts back on)

Updating goal mid-session

Send a prompt_update message to steer the agent while it works:

code

{ "type": "prompt_update", "prompt": "Focus on the authentication module" }

Injected into the system prompt before each LLM call. The agent remembers the updated goal for all subsequent turns.

Events used

Event	Handler	Purpose
`on_complete`	ulw_keep_working	Start next turn if turns remain
`before_iteration`	poll_prompt_update	Check for goal updates from frontend
`before_llm`	inject_ulw_prompt	Inject current goal into system prompt

subagents

ui_stream

ulw

Usage

Triggering ULW Mode

From the web UI

Programmatically

Turn Checkpoints

Updating goal mid-session

Events used

Enjoying ConnectOnion?