As discuss here, i suggests that we can:

1. Come up with an Agent abstraction (with every necessary method for Devin) where everyone agrees on
2. Then we can implement different Agent under agenthub/ so we can leverage our potential large pool of contributors to experiment and explore different ideas
3. While frontend and backend folks can stick with the same Agent abstraction, and when MVP implementation is completed, the same front- and back-end system can easily switch between multiple different Agent implementations

I took the first (not perfect) step towards this goal - I draft the first version of Agent abstraction, merge @rbren's agent under research/langchains_agent (pls let me know if you prefer a different name! :P ), subclass the general Agent abstraction, and make the original agents/main.py more general (you can just change a command line argument to switch between completely different agent implementation!)

And i made sure the code works now for @rbren's agent:

research/langchains_agent/build-and-run.sh

produces outputs:

STEP 0
run {'command': 'ls'}
---
output {'output': ''}
==============
STEP 1
write {'path': 'hello_world.sh', 'contents': "#!/bin/bash\necho 'hello world'"}
---
output {'output': ''}
==============
STEP 2
run {'command': 'chmod +x hello_world.sh'}
---
output {'output': ''}
==============
STEP 3
run {'command': './hello_world.sh'}
---
output {'output': 'hello world\n'}
==============
STEP 4
Done!

TODOs after this PR is merged:

• Make
• @rbren Is research/langchains_agent/regression supposed to test whether an agent can complete the tasks? If so, I'd suggest maybe we can move it into a sub-folder of evaluation and add a TODO that writes actual test cases these to test any Agent implementation against these regression tests - what do you think?
• Remove the Dockerfile for langchains_agent once Minimal Docker Sandbox with GPT-3.5 Execution Example #48 is merged (no big difference) and update build-and-run.sh correspondingly
• Merge requirements.txt with the whole project requirements.txt (or maybe not since these will be installed inside the docker container?)