You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Add Agent Teams section for keeping the AAE loop alive
When the executor agent stops and asks for input despite
instructions to loop forever, a manager/executor team setup
ensures continuous autonomous operation.
Copy file name to clipboardExpand all lines: README.md
+37Lines changed: 37 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -18,6 +18,43 @@ Agentic Algorithm Engineering takes this methodology and hands the cycle to an a
18
18
19
19
The agent must run in a mode that does not require user input, since the AE cycle is designed to run autonomously and indefinitely. If you are using Claude Code, launch it with `--dangerously-skip-permissions` (byPassPermissions mode). For other agent frameworks, ensure that all tool calls (file edits, shell commands, git operations) are auto-approved so the loop can proceed without interruption.
20
20
21
+
### Keeping the Loop Alive with Agent Teams
22
+
23
+
In practice, even with `--dangerously-skip-permissions`, the agent may eventually stop and ask for user input instead of continuing the loop. To solve this, use Claude Code's **Agent Teams** feature to set up a manager/executor pair that keeps the loop running autonomously.
24
+
25
+
**How it works:** A manager agent monitors an executor agent that runs the AAE program. Whenever the executor pauses and asks for input, the manager tells it to continue.
26
+
27
+
**Setup:**
28
+
29
+
1. Enable agent teams by setting the environment variable:
30
+
```bash
31
+
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
32
+
```
33
+
Or add it to your `settings.json`:
34
+
```json
35
+
{
36
+
"env": {
37
+
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
38
+
}
39
+
}
40
+
```
41
+
42
+
2. Launch Claude Code with `--dangerously-skip-permissions` and instruct it to create a two-agent team:
43
+
-**Manager**: monitors the executor and, whenever it stops or asks for input, sends it a message to continue the AAE loop.
44
+
-**Executor**: runs the AAE program (`AAE.md`) in the target project directory, performing the full experiment cycle.
45
+
46
+
Example prompt for the manager:
47
+
```
48
+
Create an agent team with one executor teammate. The executor should run
49
+
the AAE program defined in AAE.md on this project. Your job as manager is
50
+
to monitor the executor. Whenever it stops, pauses, or asks for input,
51
+
tell it to continue the experiment loop. Never let it idle.
52
+
```
53
+
54
+
3. The manager uses tmux split panes (or in-process mode with `Shift+Down`) to monitor the executor and send follow-up messages when needed.
55
+
56
+
This orchestration ensures the AE cycle runs truly unattended, even across context window boundaries or when the executor would otherwise break out of the loop.
57
+
21
58
## Author
22
59
23
60
Christian Schulz, [Algorithm Engineering Group](https://ae.ifi.uni-heidelberg.de/), Heidelberg University
0 commit comments