Debugging AI Agents with fork()

Our users working on agentic AI tell us that their workflows are less like a “neat flowchart” and more like a “non-deterministic trajectory.” Agents often choose the next step on the fly, from a list of possible options. Many different (but correct) execution paths are possible. The paths themselves tend to be long-running, punctuated by expensive LLM calls. So what happens when the agent stumbles onto a rare error case? Trying to “get it to happen again” may be expensive and time-consuming. Some of our users told us that they started to use the fork() feature for these cases. DBOS durable execution saves the entire trajectory. Fork can then be used to quickly re-execute a past run, from any chosen step, with a new version of the code. This way, you can quickly resolve production issues, re-test specific scenarios and develop faster. Check out this video tutorial to see how it works with one of our demo apps. Would love to hear what you think!