In the Agentic Batch Changes Beta, steps utilizing inner loop coding agents like Claude Code and Codex now stream thinking and progress in real time.
Early on, when building out Agentic Batch Changes, we decided to give the outer-loop agent that orchestrates your Batch Change effectively two tools to make code changes:
- A tool to run deterministic scripts: basically, a base container image and a bash script.
- A tool to run a coding agent by providing a prompt to an inner-loop agent to hand off work.
At first, we weren't sure if our coordination agent would see the benefit of using deterministic scripts over the simple option of using a coding agent. Coding agents can do everything now!
After letting the outer-loop agent rip through some of the most complicated problems we could come up with on the spot, though, we realized that the agent was making much smarter decisions than we could've expected, using both tools:
It would write scripts for the parts that were easy to automate because it was much easier for the outer-loop agent to predict the exact outcome. That gave the outer-loop agent more control and made things much more effective. Instead of having 500 inner-loop coding agents each solve the same dependency-installation problem in 500 different repositories, the outer-loop agent would use the knowledge it gained from studying the problem and the codebase to automate as much as possible with a script.
Then, if, during execution, it learned that there were edge cases that were hard or even impossible to script against, it would start involving coding agents. And when one coding agent goes off track, the orchestration agent would be able to steer it and apply that steering and learning to the coding agents across all other workspaces as well.
With those patterns, it was often able to make 500 changes with one outer-loop agent thread and zero inner-loop coding agents involved. And then, when they would truly make a difference, it could pull them in.
When it deployed inner-loop coding agents, following along was difficult in our first beta. The agent CLI tools we support emit beautifully structured JSON, but that output is not easy for humans to grasp.
So, for this week's update, we have:
- Improved our caching so that logs are retained when the agent continues iterating on workspaces. Logs no longer disappear from the UI when the agent makes a change for which we already have a fully cached result.
- Implemented full log streaming in the Agentic Batch Changes UI so you can always check the status of execution and guide the outer-loop agent if anything doesn't look right.
- Built human-readable agent output for the inner-loop agents. If you're keen, you can now follow exactly what the agents are currently doing and, again, provide additional guidance as needed.
The Agentic Batch Changes Beta is now available upon request: talk to your account team or schedule a demo to get started.