# ExternalMultiAgentEnv dynamics

**URL:** <https://discuss.ray.io/t/externalmultiagentenv-dynamics/21485>\
**Category:** RLlib\
**Created:** [January 14, 2025, 2:46pm UTC](https://discuss.ray.io/t/externalmultiagentenv-dynamics/21485 "2025-01-14T14:46:19Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![francesca](https://sea2.discourse-cdn.com/flex020/user_avatar/discuss.ray.io/francesca/32/7536_2.png) [@francesca](https://discuss.ray.io/u/francesca)\
**Post date:** [January 14, 2025, 2:46pm UTC](https://discuss.ray.io/t/externalmultiagentenv-dynamics/21485/1 "2025-01-14T14:46:19Z")

</div>

**How severe does this issue affect your experience of using Ray?**

- High: It blocks me to complete my task.

Hello, I’m currently using RLlib to manage a hierarchical system of agents. In this step I need to collect observation from an external source and use this data in the training.  
I’ve implemented a custom `ExternalMultiAgentEnv`, with all the necessary functions like `run()`, `get_action()`, `start_episode()`, `end_episode()` and `log_returns()`.  
In particular, I interact with the environmet trough an external script. In a call of the script I need to start the episode, get the action, apply it and in the following call I can compute the reward so log the returns and end the episode.

In the following scheme, let a for iteration be a call of the external script. This shows what I need to implement.

```auto
for i in range( episode_number):
    ... receive new data (needed to compute the reward)
    if i > 0:
         rewards = env.compute_cost(env._data)
         env.log_returns(env.episode_id,rewards,env._get_info())
    
         # end episode
         env.end_episode(env.episode_id, env._state)
        

    env.start_episode()
    env.get_action()
    ... apply the action outside the script
      

```

I tried to use `trainer.train()` as the documentations suggest, but doing like this I cannot separate the `get_action` from the `log_returns` in two successive calls. In fact it requires a full iteration to be completed inside a call.

I need to know how to implement this and still train the system of agents.