docs: add Strandslator design doc#2790
Conversation
Proposes Strandslator, an agent-driven system for translating Strands features across languages with a focus on producing reviewable PRs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
size/s
area-language
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
|
||
| ## Experiments | ||
|
|
||
| I ran early experiments translating the Anthropic model provider from TypeScript to Python, erasing existing traces from the Python SDK to set realistic conditions. With Opus 4.8 and above, translation quality is strong. The takeaway is that translation itself is not the bottleneck; what's missing is the review infrastructure, which is what shifted focus toward the artifacts and report design in this proposal. |
There was a problem hiding this comment.
Do we have objective data we can cite for this?
There was a problem hiding this comment.
I have my sample experiments but not something super extensive. This was due to timing though. Each translation workflow I kicked off ran on average 20 minutes and so I could only complete so much. I did however run an experiment with the Bedrock KB port which is committed in main now. I'll add the findings of that experiment to this doc. Overall it worked well. There were a few hiccups though because of general inconsistencies across the Python SDK, which understandably confused the agents.
| @@ -0,0 +1,100 @@ | |||
| # Strandslator | |||
There was a problem hiding this comment.
Overall I agree with the doc but I am not seeing an actionable plan that we can start implementing. From my perspective, we need the MVP to get this off the ground, then we can fill in missing pieces and improve.
There was a problem hiding this comment.
I'll add a more detailed section on this. But in short, getting something setup will be straightforward. We can follow closely what we already do with the GitHub agents. The tricky part of this system is just making sure we are consistently using it so that the code generation itself is consistent. As stated in the doc, Python will continue to be the outlier because it utilizes some of our outdated design principles.
2 participants
Summary