Interesting deployment model. The comparison to Vercel for agents is apt - the hard part isn't running the agent, it's the deployment + monitoring + retry infrastructure around it.

One question: how do you handle the handoff between filesystem state and external API state? For example, if an agent is mid-workflow when it modifies a local file but then needs to call an external API that fails - the rollback semantics get complicated fast.

For B2B automation use cases this is where most agent deployments break down. The agent does 80% of a task (enrich lead, draft email, update CRM) but when step 3 fails, nothing has a record of what happened in steps 1-2. The workflow becomes orphaned.

Does Terminal Use have any primitives for workflow checkpointing or idempotent retries?